add async hydration performance fixes

[elithrar]

This improves async hydration and prefetch performance while preserving safe clone/auth behavior.

Async repo setup and directory prefetch could do avoidable git, snapshot, and hydration-queue work. Cached reads also reopened hydrated blobs repeatedly, but naive file-handle caching risks stale reads when overlays or snapshot generations change.

• Add async repo prepare/mount flow with gated FUSE readiness and prepared-gitdir safeguards.
• Improve clone and fetch setup: skip redundant fresh async branch fetches, normalize equivalent branch refs, validate ambient credentials before clone, and use --no-tags for clone/fetch/refresh.
• Improve async hydration: dedupe queued prefetches, upgrade queued priority, re-check cache before fetch, wake workers for queued backlog, and batch snapshot size backfills.
• Improve tree indexing: stream ls-tree and cat-file --batch-check output, preserving GIT_NO_LAZY_FETCH=1 for size resolution.
• Batch directory prefetch enqueue by carrying base object and size metadata through ReaddirTyped instead of doing per-entry snapshot lookups.
• Cap speculative prefetch for very large directories at 256 prioritized file tasks.
• Cache hydrated base-file reads per FUSE file handle, with generation checks and pre/post mutation invalidation for writes, truncate, mtime updates, unlink, and rename.
• Add regression coverage for prefetch batching/capping, cached base reads, generation changes, and overlapping write invalidation.

Code path	Benchmark / signal	Before	After	Gain	Notes
Hydrator queued prefetch worker wake/ramp	BenchmarkQueuedPrefetchWorkerRamp median	42.35 ms/op	11.84 ms/op	72.0% faster	Workers now wake for queued backlog instead of waiting on slower ramp behavior.
Hydrator cached queued prefetch	BenchmarkQueuedCachedPrefetch median	82.42 ms/op	62.81 ms/op	23.8% faster	Cache re-checks avoid redundant fetches; BlobToCache calls dropped from 512/op to 0/op.
Git tree indexing	synthetic BenchmarkBuildTreeIndex median	99.18 ms/op	84.17 ms/op	15.1% faster	Streams ls-tree and cat-file --batch-check; memory dropped from 6.05 MB/op to 4.42 MB/op.
Hydrator duplicate queued prefetch	BenchmarkQueuedDuplicatePrefetch median	50.37 ms/op	48.95 ms/op	2.8% faster	Queue dedupe and priority upgrade keep duplicate prefetch work bounded.
Hydrator explicit duplicate reads	BenchmarkAsyncHydrationDuplicateReads median	47.25 ms/op	47.74 ms/op	1.0% slower	Effectively flat; left as a future optimization target rather than adding riskier waiter refactors.
Latest branch spot-check	targeted hydrator benchmarks	n/a	duplicate 35.23 ms/op, cached 48.70 ms/op, ramp 11.84 ms/op	n/a	Re-run after follow-up prefetch/read changes and merge from main.

• Validation: go build ./cmd/artifact-fs; go vet ./...; go test ./...; go test -race ./internal/hydrator; go test -race ./internal/fusefs; git diff --check.