Implements rate-limited parallel fetching of version information when solving.
Also does speculative pre-fetching of version information of dependencies of the newest versions of each package.

In informal benchmarks of pub get for a package whose only direct dependency is package:test the resolution time goes from 7.8 seconds to 1.5 seconds.

Also adds package:pedantic to our direct runtime dependencies

Do not retry on OS specific error codes for `SocketException`.
These OS error codes are not documented in the code, and they
may change behavior depending on OS and architecture.

Furthermore, many seemingly persistent errors such as DNS lookup
failures should be retried.

Finally, if we retry a `SocketException` unnecessarily this is of
little impact. It means that if the user doesn't have a network
connection, or tries to use a `PUB_HOSTED_URL` pointing to a
non-existent pub server, then it will take a bit longer before
the user is informed of the problem.

Testing `pub` without a snapshot is slow because many tests
invokes `pub` as a commandline tool. Supplying the snapshot
using a fixed filename is risky as it's easy to forget to
delete/rebuild the snapshot causing test results to be
out-of-date.

By using an environment variable to specify the name of the
`pub` snapshot, we can create a random temporary file for
each test run. This minimizes the risk that we accidentally
run tests with an outdated snapshot of `pub`.

This fixes a number of tests that failed on windows, and also two real bug fixes:

- `pub global install` now prints
  precompiling package:binary instead of precompiling package:bin\binary on windows
- #2122
 This is done by always using / for relative paths in pubspec lock path-dependencies

Previously analyzer would report these but travis would pass anyway.

Instead of writing a language version based on the current SDK.
This causes the fallback to using the latest version to happen
at runtime instead of at `pub get` time.

Use `%LOCALAPPDATA%` for `.pub-cache` on windows if `.pub-cache` doesn't already existing in  `%APPDATA%`.

Using `%LOCALAPPDATA%` is better because it's not copied when users roam between devices.

- always_declare_return_types
- annotate_overrides
- curly_braces_in_flow_control_structures
- prefer_for_elements_to_map_fromiterable
- prefer_if_null_operators
- prefer_single_quotes
- prefer_spread_collections
- unawaited_futures
- use_function_type_syntax_for_parameters

Rename a method starting with `get` to use a more accurately descriptive
verb.

Tighten the types on some methods highlighted by the
`always_declare_return_types` lint by making them generic instead of
returning `dynamic`.

Use a conditional element in a List literal instead of conditionally
calling `.insert`. Change an implementation comment into an expanded doc
comment for handling the no-unknown-keyword argument.

Use a spread element instead of `..addAll`.

A Mock is an object that can record and verify interaction during a
test, these classes are not mocks.

The `.add` method returns a boolean indicating whether then value was
not already contained in the set. Find a few places where both are used
unnecessarily and refactor to only call `.add` and use the returned
boolean.

The caller in `CacheListCommand.run` does not appear capable of handling
`null` values. Hence, filtering it out seems reasonable in this case.

Fix 2 cases of `prefer_final_fields`. Fix some unnecessary_lambda
surfaced by making one field final.

Add a missing return on an argument to `List.map`.

Change the argument type on some utility methods to allow a function
returning `FutureOr` to avoid hints about missing returns on methods
expected to return a `Future`.

And only precompile the single desired executable on pub run.

`pub run` will avoid printing "precompiling executable..." if we don't have a terminal attached.
That allows the output of `pub run` to be piped but still explains the user where time is going on first run of eg. pub run test in the terminal.

This does not change pub global behavior. We probably want to continue recompiling all executables on pub global activate.

`pub get` still allows --precompile for those who prefer the old behavior.

* Warn about old flutter plugin registration format

Remove unused Map `availableVersions`

Language versioning

* Added utility for faster local testing

This caused ordering of keys in `Map` parsed from YAML to maintain
the ordering from the YAML file.

Acoording to `doc/solver.md` pub will choose the first solution it
encounters during dependency resolution. However, the algorithm will
choose the latest version of the package with fewest versions available,
when picking the next step. Thus, this may cause pub to return a
different solution. But given that the graph of packages available is
not stable (as new versions can be published), the solutions cannot be
expected to be stable over time. Hence, changing this ordering should be
of little impact, besides reordering of messages in test cases.

Add travis badge

* Fix to show proper error message when git is not installed

* Added test case for when git is not installed

Fixes #2204

The phrase "doesn't depend on" can be interpreted as "doesn't use" and
considered an unnecessary import. Rephrase to be more clear that the
message is discussing the dependencies in pubspec.yaml.