Split submit_request from handle_response; add async submit + get_response

[JadenFiotto-Kaufman]

Summary

Refactor `RemoteBackend` so a subclass (or an async caller that wants streaming control) can submit a request without paying for `handle_response`'s side effects, and can use `httpx.AsyncClient` instead of the sync path.

Motivated by the workbench SSE endpoint flow: the FastAPI route wants to `async yield` the initial `ResponseModel` to the browser before anything touches the terminal-display logic or raises on error, and it wants to run the POST inside the event loop without a threadpool.

Changes

• `submit_request` no longer calls `handle_response`. It's now a pure POST that returns the initial `ResponseModel` (still records `self.job_id` as natural bookkeeping). Callers are responsible for dispatching the response when they're ready.
• Add `async_submit_request` — mirrors `submit_request` via `httpx.AsyncClient`.
• Add `async_get_response` — mirrors `get_response` for symmetry in the non-blocking poll path.
• Factor the shared `status_code`/`ResponseModel` parsing into a small `_parse_submit_response` helper used by both submit variants.
• Update the three internal callers (`blocking_request`, `async_request`, `non_blocking_request`) to call `handle_response` explicitly on the initial response, preserving today's behaviour. `async_request` now also uses `async_submit_request` instead of the sync one.

Test plan

• Unit / integration tests in nnsight pass on `dev`.
• Blocking remote trace (existing `with model.trace(..., remote=True)`) still shows the status display and completes.
• Async remote trace (`tracer.asynchronous=True`) still completes end-to-end; no thread blocking on the POST.
• Non-blocking submit + poll round-trip still works (submit → job_id → poll to COMPLETED).
• Workbench SSE endpoint (companion PR) exercises the new `async_submit_request` end-to-end against NDIF — verified locally against gpt2 on api.ndif.us.

Companion

Workbench side: ndif-team/workbench#112 — its `StreamingRemoteBackend` now calls `super().async_submit_request()` instead of duplicating the POST.