Add TRYBUILD_NO_CFG to opt out of cfg injection and share parent target dir

[TarunvirBains]

Adds an opt-in TRYBUILD_NO_CFG env var. When set:

• rustflags::toml omits --cfg trybuild from the rustflags passed to spawned cargo.
• cargo::cargo_target_dir returns no CARGO_TARGET_DIR override, so the parent's value (or absence) flows through.

Both halves are needed together. The cfg perturbs the rustc fingerprint, and the target-dir override means cargo never gets the chance to compare fingerprints anyway — either alone leaves the spawned cargo recompiling from cold.

When unset, behavior is unchanged. Filed as #328.

Verified end-to-end with a smoke harness:

$ cargo clean && cargo test
$ ls target/tests/trybuild/
debug/  trybuild-smoke/  x86_64-unknown-linux-gnu/

$ cargo clean && TRYBUILD_NO_CFG=1 cargo test
$ ls target/tests/trybuild/
trybuild-smoke/
$ ls target/debug/incremental/
trybuild_smoke-...

Baseline: spawned cargo writes artifacts under target/tests/trybuild/{debug,x86_64-...}. With TRYBUILD_NO_CFG=1, those subdirs are absent and the spawned cargo's incremental output lands under target/debug/ instead.

Trybuild's existing tests pass (26/26) with no test changes — they're normalization fixtures that don't spawn cargo. I considered adding a unit test that asserts rustflags::toml() returns different values based on the env var, but env-mutation tests are racy without serialization; let me know if you'd prefer it added and which approach you'd want.

Related: #305 (narrow rustflags carve-out for -C instrument-coverage), #106 / #283 (same friction from different angles).

[TarunvirBains]

Closing in favor of a different design direction.

The shared-target-dir approach in this PR works for cache reuse but introduces build-lock contention between the parent workspace and trybuild's spawned cargo. A cleaner approach is to keep trybuild's target dir isolated and seed it from the parent's artifacts (reflink / hardlink / copy fallback), preserving trybuild's isolation invariant while still getting the cache reuse.

Will follow up with a new PR. Issue #328 stays open as the problem statement.