fix: HuggingFace backend mot.usage always None without telemetry enabled

[planetf1]

Bug

ModelOutputThunk.usage is always None for the HuggingFace backend unless telemetry tracing or MELLEA_METRICS_ENABLED=true is set.

Root cause

Introduced in #653 (token metrics hooks refactor). In huggingface.py::_post_process_async, token count extraction was gated behind (span is not None or metrics_enabled), so mot.usage is never populated in plain runs.

Fix

Remove the telemetry guard from token count extraction — usage is a standard mot field, not a telemetry concern. Telemetry reporting further down still checks span/metrics_enabled as before.

# Before
if (span is not None or metrics_enabled) and isinstance(hf_output, GenerateDecoderOnlyOutput):
    ...

# After
if isinstance(hf_output, GenerateDecoderOnlyOutput):
    ...

Impact

test/backends/test_huggingface.py::test_async_avalue fails asserting mot1.usage is not None.

[planetf1]

While auditing usage consistency across backends as part of this fix, the in-process vLLM backend was found to never populate mot.usage either. Tracked separately in #696.