[3008.x] Defer OpenTelemetry imports in salt.utils.tracing / salt.utils.metrics (#69855)

[dwoz]

What does this PR do?

Fixes #69855.

salt.utils.tracing and salt.utils.metrics unconditionally imported the entire OpenTelemetry SDK at module load, even though tracing.enabled and metrics.enabled both default to false. Because both modules are imported transitively by every daemon entry point (salt.master, salt.minion, salt.channel.client, salt.channel.server, salt.utils.event, salt.netapi.rest_cherrypy.app, salt.utils.reactor), a ~15-process salt-master container was paying ~15 MB per subsystem per process — roughly 450 MB total — for functionality nobody was using.

This landed on 3008.x and shows up on the nightly stress rig as a +300 MB baseline / +430 MB peak container RSS shift versus 3006.x under otherwise-identical workload:

	3006.x	3008.x	Δ
Master container RSS, mean	1352 MB	1649 MB	+297 MB
...median	1366 MB	1712 MB	+346 MB
...p90	1529 MB	1938 MB	+409 MB
...p99	1650 MB	2065 MB	+415 MB
...peak	1683 MB	2112 MB	+429 MB

(3008.x nightly: https://github.com/saltstack/salt/actions/runs/29970269739. 3006.x nightly: https://github.com/saltstack/salt/actions/runs/28280683908.)

Each ~300 MB salt engine child inherits the same overhead (engines run as separate Python processes that re-import salt.master / salt.minion).

Change

Move the OTel imports into _load_otel() helpers invoked only after is_enabled() returns True. is_enabled() short-circuits on _cached_opts being None or enabled being false, so the probe (and therefore the OTel import) doesn't fire on the default disabled path. Public API is preserved; SpanKind remains a stub always and is translated to the real trace.SpanKind enum inside start_span() when tracing is on.

Per-process measurement of the imports removed from module top (Python 3.10 slim):

before:                        14804 KB
after tracing.py imports:      30036 KB

So each salt process trims by ~15 MB (tracing) + ~15 MB (metrics) = ~30 MB. Across the master container's ~15 Python processes that's ~450 MB reclaimed on a stack that has tracing/metrics disabled.

What issues does this PR fix or reference?

Fixes: #69855

Merge requirements satisfied?

• Changelog: changelog/69855.fixed.md
• Tests written / updated:
  • tests/pytests/unit/utils/test_tracing.py — new subprocess tests:
    • test_import_does_not_load_opentelemetry — importing salt.utils.tracing (and calling the disabled-path public API) must not put any opentelemetry.* entries in sys.modules.
    • test_enabling_tracing_loads_opentelemetry_lazily — the mirror: configure({..., enabled: True}) does trigger the import and a real span comes back.
    • test_master_and_minion_imports_do_not_load_opentelemetry — end-to-end guard: importing the whole daemon import chain (salt.master, salt.minion, salt.channel.*, salt.utils.event, salt.utils.tracing) leaves opentelemetry unimported. Catches any future eager top-level import elsewhere in the chain.
  • tests/pytests/unit/utils/test_metrics.py — matching pair: test_import_does_not_load_opentelemetry and test_enabling_metrics_loads_opentelemetry_lazily.
  • Existing test_module_works_when_opentelemetry_missing tests updated to reflect the new tri-state _OTEL_AVAILABLE (None until first probe, then True/False).
• pre-commit clean on all changed files.