feat(proxy): rescue tool calls emitted inside reasoning_content (#1)

[Hrt-Htk]

Closes #1.

What

SSEChunkLogger now rescues tool calls the chat model emits as XML
(<tool_call> / <function=…>) inside reasoning_content instead of the
tool_calls delta. Previously llama-server didn't recognize them, the turn ended
finish_reason=stop, and the client never ran the tool (the "stuck thinking"
pattern seen in the logs).

The proxy now transforms the forwarded SSE stream:

• detects tool-call XML in reasoning_content deltas across chunk/event
  boundaries (stateful capture + partial-marker holdback);
• strips the XML from forwarded reasoning while keeping genuine prose;
• re-emits it as a proper tool_calls SSE delta (args JSON-encoded, type-coerced);
• rewrites a trailing finish_reason=stop to tool_calls when a call was rescued;
• handles the no-wrapper and bare/empty <tool_call> variants and multiple calls.

Logging still observes the original deltas; only outbound bytes change.

Also fixes a latent stream-truncation bug surfaced while building this: readany()
now loops until a complete SSE event is available (or true EOF), so it never
returns b"" mid-stream — the forward loop treats empty as EOF and would
otherwise truncate responses whenever an upstream chunk split an event.

Tests

Adds tests/test_toolcall_rescue.py (9 stdlib-unittest cases): passthrough,
split-across-chunks, sub-event 13-byte chunk boundaries, prose-around-call,
no-wrapper, bare-empty, arg coercion, multiple calls per turn.

Verified offline by replaying a real captured SSE stream at several chunk sizes,
and live end-to-end against the chat proxy (normal streaming intact; no false
rescue).