⚡ (patch) De-coroutine pure-forwarder functions in the enumerator

[kammce]

Summary

Several endpoint_io/enumerator functions never actually suspend, or only tail-forward another future with nothing local borrowed across the await — each one was still paying for a full coroutine ramp/resume/destroy triple (and, where the failure path is real, an EH unwind table entry) for no benefit over a plain function:

• size_eio::driver_read/driver_write — pure synchronous bookkeeping (a fixed 0, or a length computation), never suspends at all. Converted to plain functions returning an already-complete future.
• enumerator_eio::driver_read — a pure tail call to m_ctrl_ep->read() with a matching return type and nothing borrowed across the await. Converted to return m_ctrl_ep->read(...);.
• send_error_to_host — was a coroutine only because m_retry_counter += 1 ran after the await. Reordered ahead of it (verified against every use of m_retry_counter in run()'s loop: a throwing stall() unwinds run() entirely either way, so the increment's exact position relative to a throwing stall is unobservable), then converted to a tail call.

enumerator_eio::driver_write is intentionally left as a real coroutine — its return value (p_buffer.length()) does not match what write() itself returns (see the existing TODO(#99) about actually limiting/reporting bytes written), so it isn't a pure forwarder.

Also left a TODO(#3) on write_string_view, pointing at #3: its scratch locals are borrowed across its own await, which is what blocks it from collapsing the same way — filed as a separate, smaller follow-up rather than folded in here.

Measured end-to-end on the usb demo (stm32f103zg, clang 20, MinSizeRel): 64,757 → 63,309 bytes of flash (−1,448 B).

Test plan

• test_enumerator passes
• Full downstream rebuild (libhal-arm-mcu, demos) succeeds
• usb demo binary flash size measured before/after on identical toolchain/profile