perf: optimize craft get with repo/tree caching and parallel resolution

[erdemtuna]

Summary

Optimizes craft get performance with three complementary changes that together yield an estimated 7-10× speedup for typical multi-dependency workloads.

Changes

1. Repository object caching (internal/fetch/gogit.go)

• GoGitFetcher now caches *git.Repository objects in memory
• Previously, every ResolveRef, ReadFiles, and ListTree call triggered a full ensureRepo() cycle: acquire file lock → open repo → git fetch → return → release lock
• For a single dependency, this happened 7-9 times; now it happens once
• Includes double-checked locking for thread safety and memory clone deduplication

2. ListTree result caching (internal/fetch/gogit.go)

• ListTree() results are cached by (url, commitSHA) key
• Eliminates the redundant tree walk between resolution phase (integrity computation) and install phase (file collection)
• Returns defensive copies to prevent caller mutation of cached data

3. Parallel dependency resolution (internal/resolve/resolver.go)

• Phase 4 resolveOne() loop now uses errgroup.Group for concurrent execution
• Each resolveOne() call is independent (different dep, read-only options, no shared mutable state)
• Per-repo file locks in the fetcher layer protect against concurrent cache corruption
• Added golang.org/x/sync dependency

Performance Impact

Metric	Before	After
ensureRepo() calls per dep	7-9 (each with lock+open+fetch)	1
Dep resolution	Sequential	Parallel via errgroup
ListTree per dep	2× (resolution + install)	1× (cached)
Estimated speedup (5 deps)	baseline	7-10×

Testing

• All existing tests pass
• Race detector clean (go test -race ./...)
• Pre-commit and pre-push hooks pass (vet, lint, tests)