perf(json-rpc): skip re-parse when returning owned Box<RawValue>

[mfw78]

try_deserialize_ok calls serde_json::from_str::<Box<RawValue>> even when J is already Box<RawValue>, so the raw passthrough path (Provider::raw_request, RpcClient::request/batch with Box<RawValue>) copies and re-validates a payload it already owns on every response.

When J and T are both Box<RawValue>, return the owned payload directly behind a TypeId guard, matching the transmute_copy specialization in IntoBoxTransport::into_box_transport. The generic deserialize path is unchanged.

This adds a 'static bound to the (public) try_deserialize_ok, needed for TypeId. Both in-tree callers pass Box<RawValue>; only external callers passing a borrowed J would be affected.

[mattsse]

this seems okay, we only use it for raw_request_dyn and there we already enforce 'static owned I believe so the additional 'static is okay here.

wdyt @DaniPopes

[mfw78]

The failing docs job is an upstream rustc regression, not this PR. The 2026-07-02 nightly ICEs in rustc_metadata (get_item_attrs decoder assertion) when rustdoc inlines a pub extern crate re-export, which here is prost::alloc. It is tracked in rust-lang/rust#158686 and triggered by rust-lang/rust#158569. It reproduces on main without this change (the base commit ICEs identically), and every other check passes. The job should clear on a nightly that carries the fix, or on a re-run once it lands.