fix(diagnostics): never send diagnostics from test environments

[naji247]

Summary

MCPCat's internal SDK diagnostics exist to catch broken installs — they should never fire from a test run, neither ours nor yours. Before this change, diagnostics were enabled by default with no test-environment awareness, so any test suite that called track() could ship real OTLP metadata (host/OS/runtime info) to the diagnostics collector. This PR makes staying silent in tests the default, enforced at two layers:

• SDK-level guard (protects every consumer): diagnostics now auto-disable whenever a test environment is detected — PYTEST_CURRENT_TEST or PYTEST_VERSION set. No configuration needed; your CI sends nothing. Anyone who genuinely needs diagnostics active in a test context can opt back in explicitly with DISABLE_DIAGNOSTICS=false.
• Repo-level guard (protects this repo's CI): a top-level tests/conftest.py sets DISABLE_DIAGNOSTICS=true before any test runs, so no test ordering or future change to the detection logic can leak traffic from our own suite. Diagnostics-specific tests opt back in per-file with mocked HTTP, keeping them network-safe.

Also tightens DISABLE_DIAGNOSTICS parsing: explicit falsy values (false/0/no/off) are now a deliberate opt-in that overrides test detection, and whitespace-only values are treated as unset. The disable_diagnostics docstring documents the new behavior.

Counterpart of the TypeScript fix: agentcathq/agentcat-typescript-sdk#44.

Test plan

• Regression test written first, confirmed failing (RED), then passing (GREEN). tests/test_diagnostics_test_env.py::test_track_does_not_enable_diagnostics_in_pytest — before the SDK guard, track() under pytest latched diagnostics on (assert True is False); after the guard it stays disabled and flush_diagnostics() posts nothing.
• New opt-out cases per detection signal (tests/test_diagnostics_optout.py): enabled by default in a simulated non-test env; disabled by default when PYTEST_CURRENT_TEST set; disabled by default when PYTEST_VERSION set; DISABLE_DIAGNOSTICS=false force-enables even under pytest; whitespace-only treated as unset; falsy values force-enable.
• Full suite: uv run pytest → 479 passed, 4 skipped, 3 xfailed.
• Dead-endpoint run: DIAGNOSTICS_ENDPOINT=http://127.0.0.1:9 uv run pytest → 479 passed, 4 skipped, 3 xfailed (no hangs, no leaked traffic).
• Lint: uv run ruff check . — no new findings (new files clean; pre-existing repo findings unchanged).
• Types: uv run mypy src/mcpcat --strict — error count unchanged from baseline (130), no new errors introduced.