(For 2.12.x's eyes only.)

Use vals to cache a single instance of stateless CanBuildFrom instances.
These are cast by the existing `implicit def` to the suitable generic
type. This pattern was already used in some places -- this PR applies
it systematically across `collection.{mutable,immutable}`.

The `CanBuildFrom` instances for arrays and wrapped arrays are
cached for each primitive type, Unit, and Object. Each of these
instances is backed by a dedicated subclass of CanBuildFrom that
avoids subsequent dispatch on the `ClassTag[T]`.

[nomerge] Backport windows test fix

order matches in expected frequency order (Array, WrappedArray + associated builders, and ClassTag.newArray)
avoid extra def in BitSets
don't optimise for NoBuilder cases

Closes scala/bug#7216
Closes scala/bug#7288

two more tests for apparently-fixed issues

[nomerge] ConcatIterator.last advances

[backport] Import travis caching config and JDK install from scala-dev

use direct vals where no casting is required

Avoid allocations of reusable CanBuildFroms

The test allocates 16*16 MB, so testing with a 192 MB heap should
be safe to catch the leak.

Tested with running the test in a `while true` loop locally
  - constanly fails with 192 MB on 2.13.1
  - constantly succeeds with 192 MB on 2.13.2-bin-9ef8fc3
  - flaky with 128 MB on 2.13.2-bin-9ef8fc3

(cherry picked from commit ffcffc6)

[backport] Bump Xmx for type-tag-leak

  - Make `ImportContext` a nested class
  - Avoid the constructor of the superclass `Context` needing to
    call methods on the unitialized subclass ImportContext by
    computing `isRootContext` / `depth` externally in the
    factory method.

Rather than adding a wrapper object for each import in the session
history, just use a single wrapper preceded by the imports which
have been interspersed with a magic import to bump context depth.

Code is still ordinarily wrapped in a `$read` object.

This is a step toward 6623-like transparency.

`retronym` takes the blame for this innovation.
`adriaanm` collaborated in its commission.
`somsnytt` batted clean-up.

To be reverted.

… bug

By default, the parser tries to heal incomplete source files by inserting
missing braces. Compilation will still error out, but any subsequent
parser/type errors make more sense to the user when this healing works.

The healing uses indentation to figure out the intent of the code. Wrapped
REPL snippets aren't properly indented, and in the test case I added the
healing seems counterproductive.

This commit disables it in REPL tab completion, in line with the way that
IMain parses the wrapped code for real compilation.

  - add comments describing a problem I discovered with imports
    and a possible solution
  - Make strip wrappers from REPL output for the wrappers generated in this mode

This detail changes under -Yrepl-class-based. We prefer tests that operate
under either mode.

Revert "Enable magic imports unconditionally to survey test results"

This reverts commit b02f890.

Revert "Enable -Yrepl-class-based to survey test failures"

This reverts commit d64af1a.