perf(tune): read CPU load directly instead of collecting every signal

`machine_is_loaded` called `collect_signals()` to obtain one number. That
function also constructs an AI-workload monitor, snapshots every GPU through
NVML, and enumerates the process table twice.

Measured on this machine: 1465 ms against 0.2 ms for the platform CPU reader
alone. Roughly seven thousand times the cost, paid on every verified apply of
a load-dependent metric, purely as a precondition check before any
measurement starts.

`serve::read_cpu_utilization` carries a comment saying it deliberately avoids
`ontology::resolve::snapshot()` because resolving every domain to obtain one
number took tens of seconds per cycle. I wrote the same mistake three
functions below that warning, in the same file, yesterday.

The numbers are measured rather than asserted, which is the only reason I
know the factor is seven thousand and not seven.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>