feat(telemetry): record root cause on WinMLCLIError

[timenick]

Summary

• Record the innermost root cause of an exception chain on WinMLCLIError, so wrapped errors (e.g. ClickException around a ValueError) become analyzable instead of collapsing to ErrorType=ClickException.
• New attributes root_cause_type (for grouping) and root_cause_message (scrubbed, capped at 500 — larger than the outer 200 cap since the root cause is the diagnostic payload). Both are None when the outer exception has no deeper cause, so "no chain" is visually distinct from a real root cause.

Changes

• telemetry/utils.py: new _root_cause chain-walker (prefers __cause__ over __context__, walks to innermost, cycle-safe via id()); _format_exception_message gains a cap param; new _ROOT_CAUSE_MESSAGE_CAP = 500.
• telemetry/telemetry.py: log_error resolves the root cause and emits the two new attributes; both added to the WinMLCLIError allowlist.

Notes

• This is Part 1 (telemetry layer) of WinMLCLIError telemetry loses root cause — ClickException wrappers make errors unanalyzable #1110. Part 2 (fixing command-layer sites that drop the cause, e.g. commands/quantize.py:327) is intentionally out of scope — the plan is to use the new telemetry data to prioritize the highest-frequency offenders.
• The 500-char cap is an initial value; it can be tuned once real telemetry shows how often root-cause messages are truncated.

Part of #1110

[xieofxie]

Nice, well-scoped Part-1 change. The chain-walker is clean — __cause__-over-__context__ preference, innermost walk, and id()-based cycle guard are all correct and thoroughly tested (I ran the full telemetry suite locally: 224 passed). The cap parameterization preserves the scrub-before-cap ordering, and has_root = root is not exc gives a clean null distinction for "no chain". Splitting the command-layer from-dropping fixes into Part 2 is the right call.

One behavioral nuance on _root_cause worth considering before merge (inline) — otherwise LGTM.

[xieofxie]

Re-reviewed at eae6f92a. The __suppress_context__ point is fixed exactly as suggested: _root_cause now only falls through to __context__ when it isn't suppressed, so raise ... from None correctly surfaces no chain — matching Python's own traceback behavior and the developer's intent to hide the inner error. The docstring is updated and test_root_cause_honors_suppressed_context locks it in. Ran the affected telemetry tests locally: 66 passed. LGTM.