telemetry: add per-metric activation (NIX-453) (#1919)

## What?

Adds **per-metric activation** for telemetry via
`NIXL_TELEMETRY_METRICS` — a
comma-separated glob allowlist (POSIX `fnmatch`, e.g. `agent_tx_bytes`,
`agent_err_*`) selecting which metrics are exported.

- Resolved **once at construction** into an enum-indexed mask.
- **Gated at the producer**: `updateData()` / `addXferStats()` skip
deactivated
  metrics *before the staging queue*, so a deactivated metric costs no
mutex/append on the transfer hot path and is never counted as a drop.
The
check is a single lock-free read of the immutable mask
(`isMetricEnabled`).
- **Unset/empty = all metrics active** (backward compatible); a token
matching
  nothing is ignored with a warning.
- `getXferTelemetry()` is unaffected (reads the request handle, not the
queue).

Docs (`docs/telemetry.md`, prometheus/doca/plugin READMEs) and tests
updated in
the same PR.

## Why?

Tracked by NIX-453 (parent NIX-1423): operators need to enable/disable
individual telemetry metrics without turning telemetry off. Gating at
the source
(rather than at drain) means deactivated metrics incur zero hot-path
cost — per
review feedback from @AlexNvd.

This is **Phase 1** — event-type granularity, core-only, collision-free
with the
in-flight histogram PR #1914. **Phase 2** (NIX-1612: per-series /
histogram-group control, `*_us`/`*_total` globs) lands after #1914.

## How?

- `telemetry_event.h`: `nixl_telemetry_event_type_count` +
`nixl_telemetry_metric_mask_t`
  alias (append-only; no descriptor changes).
- `telemetry.cpp`: `resolveEnabledMetrics()` parses
`NIXL_TELEMETRY_METRICS`
(`absl::StrSplit` + `fnmatch`) into the mask;
`updateData()`/`addXferStats()`
gate on `isMetricEnabled` before staging; `flushPendingEvents()` no
longer filters.
- Tests (gtest, 33/33): producer-gating cases (subset, family glob,
unknown-token)
  on the BUFFER stream + an end-to-end Prometheus scrape case.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Added `NIXL_TELEMETRY_ENABLED_METRICS` to restrict exported telemetry
using a comma-separated POSIX-glob allowlist (selects whole metric
families).
* **Bug Fixes**
* Deactivated metrics are filtered out before buffering/staging, and the
“telemetry dropped” event is emitted only when enabled.
* **Tests**
* Added Prometheus and buffer-export allowlist tests (subset, family
globs, and unknown tokens). Added a disabled performance benchmark for
fully deactivated per-transfer stats.
* **Documentation**
* Updated telemetry and exporter docs to cover matching rules, warnings
for unmatched tokens, and default “export all” behavior.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: Efraim Eygin <eeygin@nvidia.com>