feat: respect CPU and cgroup limits in server-ng and connectors stats

[mmodzelewski]

The server-ng and connectors runtime stats endpoints reported
host-wide sysinfo numbers: a cpuset-confined or memory-capped
process showed its neighbors' CPU load and a memory total it can
never allocate, leaking host sizing on multi-tenant machines. The
legacy server already scoped these (#3615) but kept the logic
inline.

Extract that logic into a shared system_stats crate: SystemProbe
scopes total CPU usage to the allowed core set (snapshotted at
startup, before shard threads pin themselves) and memory totals
to the effective cgroup cap, falling back to host-wide values for
unconfined processes. All three binaries now probe through it.

On confined hosts the reported total_cpu_usage, total_memory and
available_memory now reflect the confinement, so dashboards and
alerts keyed to the old host-wide values will see lower numbers.

The connectors runtime also keeps its sysinfo System alive across
captures, so CPU usage reflects real deltas instead of the zero
first sample of a freshly created System on every request.

[codecov]

Codecov Report

❌ Patch coverage is 88.98678% with 25 lines in your changes missing coverage. Please review.
✅ Project coverage is 57.79%. Comparing base (a860cef) to head (a3338e2).

Files with missing lines	Patch %	Lines
core/system_stats/src/lib.rs	84.74%	14 Missing and 4 partials ⚠️
core/server/src/shard/system/stats.rs	91.04%	6 Missing ⚠️
core/server-ng/src/main.rs	0.00%	1 Missing ⚠️

Additional details and impacted files

@@              Coverage Diff              @@
##             master    #3647       +/-   ##
=============================================
- Coverage     73.42%   57.79%   -15.64%     
  Complexity      937      937               
=============================================
  Files          1285     1284        -1     
  Lines        140958   124703    -16255     
  Branches     116848   100634    -16214     
=============================================
- Hits         103503    72071    -31432     
- Misses        34228    49570    +15342     
+ Partials       3227     3062      -165     

Components	Coverage Δ
Rust Core	54.15% <88.98%> (-19.49%)	⬇️
Java SDK	62.44% <ø> (ø)
C# SDK	71.40% <ø> (-0.71%)	⬇️
Python SDK	91.65% <ø> (ø)
PHP SDK	84.29% <ø> (ø)
Node SDK	91.26% <ø> (-0.10%)	⬇️
Go SDK	42.28% <ø> (ø)

Files with missing lines	Coverage Δ
core/connectors/runtime/src/main.rs	85.71% <100.00%> (+0.05%)	⬆️
core/connectors/runtime/src/stats.rs	70.58% <100.00%> (-1.64%)	⬇️
core/connectors/sdk/src/api.rs	100.00% <ø> (ø)
core/server-ng/src/responses.rs	32.42% <100.00%> (-0.72%)	⬇️
core/server/src/main.rs	63.34% <100.00%> (+0.10%)	⬆️
core/system_stats/src/cgroup_memory.rs	86.30% <ø> (ø)
core/server-ng/src/main.rs	0.00% <0.00%> (ø)
core/server/src/shard/system/stats.rs	90.27% <91.04%> (+6.94%)	⬆️
core/system_stats/src/lib.rs	84.74% <84.74%> (ø)

... and 316 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[hubcio]

one finding lands outside the diff so it goes here: ConnectorRuntimeStats in core/connectors/sdk/src/api.rs now carries cpuset/cgroup-scoped total_cpu_usage / total_memory / available_memory but the fields have no rustdoc - legacy Stats got per-field scoping notes in #3615, worth mirroring the memory ones here (hold the cpu wording until the affinity issue below is fixed, right now it would document a value the servers don't produce). release notes should also mention the value change for server-ng stats: total/available memory become the cgroup cap on confined hosts. wire-compatible, but dashboards and SDK consumers see different numbers.

[hubcio]

all findings from the earlier review are fixed as of fb6d2f8 - verified each one: boot-time capture_allowed_cpus() snapshot runs on the main thread before any shard pinning in all three binaries, and the .get()-not-get_or_init read means a missed capture falls back to host-global instead of lazily grabbing a pinned thread's mask; System::new() swap landed everywhere including the two pre-existing sites; positive averaging test + None-branch assertion added; cgroup_scoped_memory got the calling-process doc and went private; Pid bound once. cargo test -p system_stats green, fmt + clippy clean.

two leftovers, neither blocking:

• ConnectorRuntimeStats in core/connectors/sdk/src/api.rs still has no per-field rustdoc on total_cpu_usage / total_memory / available_memory, which now report cpuset/cgroup-scoped values on confined hosts. legacy Stats in core/common/src/types/stats/mod.rs got per-field scoping notes in #3615 - worth mirroring, and the cpu wording is safe to write now that the affinity snapshot landed.
• the PR description still names scoped_total_cpu_usage / cgroup_scoped_memory as the crate's api (both are private now - the surface is SystemProbe::capture + capture_allowed_cpus) and doesn't mention the affinity boot snapshot or that stats values change for dashboards on confined hosts. worth a refresh before merge, it doubles as the release note.