feat: add permit APIs and notification_capacity to HandlerCtx (#40)

Expose backpressure-aware permit APIs on HandlerCtx that mirror
tokio's mpsc::Sender reserve surface. Users can now reserve channel
capacity before sending, with both borrowed and owned permit types.

New types: NotifyPermit, OwnedNotifyPermit
New methods: permit, permit_owned, try_permit, try_permit_owned,
permit_many, try_permit_many, notification_capacity

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

v0.6.1 - notify a stream

feat!: add consuming into_raw_value to RpcSend (#37)

* feat!: add consuming `into_raw_value` to `RpcSend`, remove `Serialize` supertrait

`RpcSend` previously required `Serialize` as a supertrait, forcing lazy
producers like iterators to use `Mutex<Option<T>>` wrappers for
interior mutability during serialization. This replaces the marker trait
with a consuming `into_raw_value(self)` method, enabling direct
ownership transfer during serialization.

The blanket impl for `T: Serialize` is preserved, so all existing code
that returns `Serialize` types from handlers continues to work. Types
that do not implement `Serialize` can now implement `RpcSend` directly
with custom serialization logic.

BREAKING CHANGE: `RpcSend` no longer has `Serialize` as a supertrait.
Code that used `T: RpcSend` and relied on the `Serialize` bound will
need to add an explicit `Serialize` bound. `HandlerCtx::notify` now
takes `T` by value instead of `&T` (callers passing `&item` still work
via the blanket impl).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: address review feedback

- Deprecate `ErrorPayload::serialize_payload` in favor of `into_raw`
- Eliminate extra serialization round in `build_response` by constructing
  the JSON envelope directly via format!, bypassing the Response struct

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: use types instead of hand-crafted JSON in response serialization

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor: generalize Response Serialize bounds to RpcSend

Response now owns the payload via Cell<Option<...>> so the Serialize
impl can consume it. Bounds are T: RpcSend, E: RpcSend instead of
T: Serialize, E: Serialize. build_response no longer needs to call
into_raw explicitly — it's handled inside the Serialize impl.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Revert "refactor: generalize Response Serialize bounds to RpcSend"

This reverts commit 1383288.

* fix: memoize parse_error response and serialization_failure payload

- parse_error() now clones from a LazyLock<Box<RawValue>> instead of
  serializing on every call
- serialization_failure() borrows a static ResponsePayload instead of
  reconstructing it per invocation

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

v0.4.1 - the one with docs that work?

v0.4.0 - the relase with otel semconv

0.3.4 - more permissive deserialization

v0.3.3 - axum ws convenience

v0.3.2 -- axum WS support

v0.3.1

- bunch of span improvements for better traces