perf: Deduplicate file hashing and parallelize globwalks (#11902)
## Summary
Optimizes `turbo run --dry` wall-clock time by up to 1.48x on large
monorepos by eliminating redundant file hashing work and removing a
serialization bottleneck in globwalk operations.
### Benchmarks
Tested across three repos of varying size:
| Repo | Packages | Before | After | Speedup |
|------|----------|--------|-------|---------|
| large | ~1000 | 5.903s | 3.999s | **1.48x** |
| medium | ~120 | 1.461s | 1.380s | 1.06x |
| small | ~6 | 0.659s | 0.693s | ~1.0x (noise) |
The improvement scales with repo size — specifically with how many tasks
share the same `(package, inputs)` combination.
### Changes
**File hash deduplication** — Multiple tasks in the same package with
identical `inputs` config (e.g. `build`, `lint`, `typecheck` all in one
package) previously each ran an independent globwalk + file hash
computation. Now tasks are grouped by `(package_path, globs,
include_default)` and each unique combination is computed once, with
results shared across tasks.
**Parallel globwalks via retry-on-EMFILE** — The previous `IoSemaphore`
(max=1) serialized all globwalk operations to prevent fd exhaustion,
making this the dominant bottleneck on large repos. This replaces the
semaphore with retry-with-exponential-backoff on `EMFILE` errors (the
same pattern Node's `graceful-fs` uses), allowing globwalks to run fully
parallel on rayon. If the OS returns "too many open files", the
operation sleeps briefly and retries — up to 10 times with exponential
backoff capped at 1s.
**Zero-copy lockfile dependency lookups** — `Lockfile::all_dependencies`
now returns `Cow<'_, HashMap<String, String>>` instead of cloning the
HashMap on every call. For pnpm (which pre-builds a dependency index),
this eliminates ~329k HashMap clones during transitive closure
resolution.
**Optimized transitive closure cache keys** — The `DashMap` resolve
cache now uses a single null-byte-separated `String` key built into a
reusable buffer, instead of allocating a `(String, String, String)`
tuple per lookup.
**HashMap importers for pnpm** — Converted pnpm's `importers` field from
`BTreeMap` to `HashMap` (with sorted serialization) for O(1) workspace
lookups during `resolve_package`.
