fix(mediorum): avoid blocked replication pipe goroutines

[raymondjacobson]

Summary

• prevent replicateFileToHost pipe writer goroutines from blocking forever when peer replication returns early
• close the request body pipe on SignedPost/HTTP/non-200 failures so the multipart writer unblocks
• add a regression test for an HTTP 500 peer that does not consume the request body

Context

In production, creator-2 had ~42k goroutines, with ~41k parked in MediorumServer.replicateFileToHost.func1 at pkg/mediorum/server/replicate.go:191. Those goroutines were blocked sending to an unbuffered errChan after failed repair/replication attempts. The pod memory graph was mostly file cache, but the openaudio heap was also inflated compared with creator-1.

The old code could return from the HTTP path before reading from errChan, and if the peer did not consume the body or the pipe copy failed, the writer goroutine could remain blocked.

Test Plan

• go test -c ./pkg/mediorum/server
• attempted go test ./pkg/mediorum/server -run TestReplicateFileToHostClosesPipeOnEarlyHTTPError -count=1, but local package TestMain requires Postgres on localhost:5454; it failed before running the test with connection refused.