wtf_transcribe: audio fingerprint cache + NFS retry

[pavanputhra]

Summary

• Replaces the Redis transcription cache in the wtf_transcribe link with a Shazam-style audio fingerprint cache (FFT spectral peaks → peak-pair hashes), kept as a per-worker in-memory dict (fingerprint_cache.py).
• Adds an NFS ESTALE/ENOENT retry loop on file:// reads so transient mount blips no longer fail the link.
• Sets cachish: true on the vfun POST so the upstream service knows the call is cache-aware.
• Adds libchromaprint1 to the runtime image and scipy to project deps (FFT).

Authorship

Code authored by @howejacob1 — opening on his behalf so review can start. He can take over the branch or push fixups directly.

Test plan

• Build image, run conserver against a chain that uses wtf_transcribe
• First call: cache miss, vfun hit, fingerprint stored
• Second call with the same audio: cache hit, no vfun call
• Simulate an NFS ESTALE: confirm retry succeeds instead of failing the link