server: scope SCCACHE_BASEDIRS to compiler requests

[tcashel]

Problem

I often have several coding agents build the same Rust project from separate Git
worktrees. Each build can identify its own checkout root, but the sccache daemon
currently captures SCCACHE_BASEDIRS when it starts. Supporting another
worktree therefore requires restarting the daemon or adding every checkout to
one global list.

This draft depends on #2813 and contains one commit on top of it. Until #2813
merges, GitHub includes both commits in this PR; view the follow-up-only
diff.
I'll rebase this branch onto main after #2813 merges.

Change

Each compiler invocation can now override the file-configured basedirs. If
SCCACHE_BASEDIRS is absent, the invocation uses the configured fallback. An
empty value disables that fallback, while a relative value fails only that
invocation. Each request gets an isolated storage view, so concurrent builds
cannot affect one another. The same rules apply with SCCACHE_CLIENT_SIDE=1.

For example, these builds can use one running daemon:

cd /work/project
SCCACHE_BASEDIRS="$PWD" cargo build

cd /tmp/project-worktree
SCCACHE_BASEDIRS="$PWD" cargo build

Only the config file defines the daemon fallback, and statistics continue to
report only that value. This change does not discover Git worktrees, expand
wildcard roots, or add a wire-protocol field.

Validation

On Apple silicon, the required formatting, clippy, and locked test suites passed.
Focused Cargo tests cover daemon reuse, both client modes, override precedence,
configured statistics, and concurrent requests.

I also ran the final branch on Ubuntu x86_64. The full locked suite and all
single-backend builds passed. NVCC 12.4 ran both CUDA system-test modes without
skipping. In a two-root canary, request-scoped basedirs added a cross-root
CUDA hit while one daemon stayed running; the device-code and PTX stages still
missed.

The repository benchmark suite also completed. In one local run,
cache_key_generation had a 31.87 µs median, compared with 31.41 µs in the
PR #2813 run.

Fixes #2595.

I used OpenAI Codex (gpt-5.6-sol, ultra reasoning) during implementation,
testing, benchmarking, and review. I'll review the full diff and CI results
myself before marking the PR ready.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 97.34513% with 21 lines in your changes missing coverage. Please review.
✅ Project coverage is 74.10%. Comparing base (8ab3926) to head (645971d).

Files with missing lines	Patch %	Lines
tests/sccache_cargo.rs	95.81%	9 Missing ⚠️
src/server.rs	80.76%	5 Missing ⚠️
src/cache/basedirs.rs	96.93%	3 Missing ⚠️
src/compiler/rust.rs	99.17%	2 Missing ⚠️
src/util.rs	98.40%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2814      +/-   ##
==========================================
+ Coverage   73.71%   74.10%   +0.38%     
==========================================
  Files          72       73       +1     
  Lines       37932    37684     -248     
==========================================
- Hits        27963    27927      -36     
+ Misses       9969     9757     -212     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[sylvestre]

please add the benchmark in a different pr