fix: report no requests figure where the engine exposes none

[outofcoffee]

The llama.cpp metrics tile's requests figure was always 0, even though the
other engine-sourced counters worked. The collector read the figure from
request_success_total for every engine family, but that metric is vLLM's:
llama.cpp's server exposes no cumulative request counter at all, so the figure
was a zero no engine ever produced.

Summary

• The statistics now carry the request count only where the engine family's
  metrics expose a cumulative request counter (vLLM's
  request_success_total); a family that exposes none (llama.cpp, today)
  omits the field rather than reporting a zero.
• TokenStats.Requests becomes optional (*int, omitempty) across the
  daemon's reply, the remote relay, and the metrics JSON output.
• The shared token block — dashboard tile, the metrics bar and table
  formats, the serve view — draws the requests: line only for a figure the
  statistics carry.
• The unit-test fixture and the two example stacks' engine configs stop
  fabricating a llamacpp:request_success_total line a real engine would
  never serve.
• The OpenAPI description and the control plane's TypeScript mirror mark the
  field as optional.

Implementation details

The absence is modelled as a pointer, not a zero the renderers hide: a
genuine zero from an engine that exposes the counter (started, nothing served
yet) still serialises, while a family that exposes none omits the field
entirely. That keeps "no figure" and "a zero figure" distinguishable in the
data, the same rule the engine-metrics spec already applies to the last-active
pair and the system stats — and it is what lets every rendering surface drop
the line without special-casing a runner.

Verified against a live llama.cpp server run with --metrics: the collector
yields the token counters with no requests field, and no surface draws a line
for it.

Fixes #212.