Added the CHASM stream component.

[moetemp]

This PR adds the CHASM stream component, with its retention, caps and producer dedup.

What changed?

• chasm/lib/stream/stream.go is the component. It holds records in batches next to the owning workflow, with a frontier, producer cursors and consumer pins. Appends, closing, truncation, cap reclaim and the read window live here. Truncation never passes an active consumer's floor. A filtered read moves only past the messages it examined, so a short page skips nothing.
• Producer dedup keeps one entry per producer id and matches on its sequence. A record can declare its identity under temporal.io/content-hash (fingerprint.go). That's a binary/plain payload with the 64 hex bytes of the SHA-256 over the converted body, taken before any codec (stream.ContentHashOf). The table compares that hash, so a codec with a per-call nonce can't turn a retry into a conflict. Only appends that name a producer read the key. A value that isn't hex falls back to the encoded fingerprint.
• reason.go: refusals a caller must act on are FailedPrecondition with a documented message prefix. The tokens are STREAM_PRODUCER_CONFLICT, STREAM_PRODUCER_STALE_SEQUENCE, STREAM_CURSOR_BELOW_FLOOR and STREAM_CLOSED. An SDK maps them to typed errors until a typed detail carries the reason on the wire.
• cursor.go: the consumer cursor lives on the reader, not the stream. A delivered range then commits with the event that records it.
• config.go holds the caps as namespace dynamic config, plus a byte and item budget for a stream a workflow owns, well under the mutable state limit. The budget counts held records, head minus base, so a stream created above zero isn't over budget at birth. messages.go caps the bytes and count one read returns.
• StreamLifecycle.max_bytes and StreamState.held_bytes give the creator a byte cap beside the record cap. Reclaim keeps the count exact. An append past the cap gets the storage-limit ResourceExhausted and reclaims nothing.
• Retention is an age on an open stream. Each batch carries its append time. The first append arms a check that reclaims older batches and re-arms on stream.retentionRecheckInterval while records remain. It never moves the floor past an active consumer. Closed streams keep the post-close deletion.
• proto/v1/{message,stream_state,tasks}.proto define the record, the batch, the state and the two side-effect tasks. A sequence of zero means the producer doesn't number. The close reason on Terminate is encoded with payload.EncodeString, so a data converter reads it.

This layer wires nothing up. The service, the commands and delivery come in #5 to #7.

Part of AI-198 (epic AI-37).

Why?

The payload lives in the component, not in History and not in a log table. Everything above this layer depends on that property, so it's worth reading before the wiring.

How did you test it?

go build ./..., make proto and make fmt are clean. The ./chasm/lib/stream/... unit tests pass, with the age, cursor, fingerprint and read-cap suites. Nothing here runs in a cluster yet. Limits: a registered workflow consumer makes max_items a lifetime quota until it deregisters, since replay re-reads every range. One entry per producer id allows one append in flight, so pipelining needs an id per lane. A query or reset of a closed run is refused once truncation takes the floor ForgetConsumer released. Every limit treats zero as the default.

• Unit Tests
• Staging
• End to End Tests

[moetemp]

Replaced by #18, #19, #20, #21, #22, #23, #24, #25, #26, #27, #28, #29, #30, #31, #32, #33, #34, #35, #36 and #37.

Same content, split into 20 PRs in the v3 series: the notification channel first, then the streaming interface, then native streams and the rest. The branch stays as a pin.