fix(fuse): close OpsNode bypass gap in FUSE_INTERRUPT decoupling

[mfreed]

Three OpsNode methods (Lookup, Getattr, Readlink) bypass FSAdapter and call Operations directly. The previous fix (4c7b4c1) wrapped FSAdapter.{Stat, ReadDir, ReadFile} with decoupleFromRequestCancel, but the bypass paths kept propagating the kernel's per-request context all the way into pgx.

Symptom: docker-FUSE stress runs hit a transient ENOENT for a recently-created/renamed file, recovering only after the 2s statCacheTTL. Pattern: kernel sends FUSE_INTERRUPT during a LOOKUP (typically from Go's SIGURG goroutine preemption under GC pressure), go-fuse cancels the request ctx, the cancellation reaches db.ResolvePath via the OpsNode bypass, pgx returns context.Canceled, resolveSynthPath translates it to "", false, statSynthFile turns the resulting ErrNotExist into a setNegative entry, and subsequent reads (including the runner's retries) hit the stale negative for the full 2s TTL.

Fix: same one-line wrap at the entry of each bypass site -- ctx = decoupleFromRequestCancel(ctx). Three lines plus doc-comment blocks explaining each method's bypass.

Validation:

• Before: ~20% of docker-FUSE 1500-iter stress runs hit the cache-bug signature (probe: "DB has row; FS recovered visibility at offset 2000ms").
• After: 0/8 attempts hit it across ~12000 total iterations. Two attempts hit DIFFERENT failure shapes (close-time EIO, rename mid-flight ENOENT) -- pre-existing different bugs, out of scope.