fix: surface aborted partial cache writers as errors, not clean EOF

[stareezy-1]

Summary

Fixes #932

Problem

A streaming partial-cache reader treats the writer-side sender dropping as clean end-of-body:

// wait for more data
if self.bytes_written.changed().await.is_err() {
    // err: sender dropped, body is finished
    // FIXME: sender could drop because of an error
    return None;
}

But the storage contract says a miss handler dropped without finish() is a failed write. When a cache-miss writer is cancelled before EOF and before finish(), the partial reader reports Ok(None), which the proxy converts to HttpTask::Done — the client receives a truncated response that appears complete.

Reproduction: start a streaming write on MemCache, read the partial body, then drop the writer without finish(). The reader returns Ok(None) (clean EOF) instead of an error.

Fix

1. PartialHit::read() returns an error when the sender drops while the state is still Partial — the writer never reached EOF, so the admission failed.
2. MemMissHandler::finish() sends the Complete marker before consuming the handler, so a successful admission still yields a clean EOF even when the final write_body had eof=false. The watch value is copied out before send_replace (which takes the watch's internal write lock) to avoid a self-deadlock with the borrow Ref.

After the fix, the reader state machine is unambiguous: Complete state = real EOF; sender drop while Partial = aborted writer (error → failed response, not truncated success).

Tests

test_partial_reader_errors_on_aborted_writer in memory.rs:

Step	Verifies
Writer writes partial data; reader consumes it	partial read works
Writer dropped without finish()/EOF	—
Reader's next read_body() returns Err	aborted writer is not clean EOF — fails without the fix
test_write_while_read (existing)	finish() after non-EOF writes still yields clean EOF

Full crate suite passes (145 tests), cargo fmt --check clean, clippy clean for the touched file.