Improve telemetry error classification to reduce 'unknown' error types

[StellaHuang95]

Problem

The classifyError() function in the telemetry error classifier had limited coverage — only 6 error categories were recognized. Many errors from manager registration (MANAGER_REGISTRATION_FAILED) were falling into the 'unknown' bucket, making it difficult to diagnose failures from telemetry data.

After tracing every code path that leads to MANAGER_REGISTRATION_FAILED telemetry (pipenv, pyenv, poetry, conda, system, and shellStartupVars registration), we identified 6 common error patterns that were unclassified.

Changes

Added 6 new error types to DiscoveryErrorType and corresponding classification logic:

New Type	What It Catches	Detection Method
tool_not_found	"Conda not found", "Python extension not found", "Poetry executable not found"	Message pattern
command_failed	Failed to run "conda ...", Failed to run poetry ..., Error spawning conda: ...	Message pattern
connection_error	PET process dies mid-request, JSON-RPC connection closed/disposed	instanceof rpc.ConnectionError
rpc_error	PET returns a JSON-RPC error response (e.g., internal error, method not found)	instanceof rpc.ResponseError
process_crash	PET restarting, failed after N restart attempts, failed to create stdio streams	Message pattern
already_registered	Duplicate manager registration (race condition)	instanceof BaseError

Every message pattern is based on exact throw new Error(...) strings found in the codebase — no speculative matching.

Ordering

The check order is designed to avoid misclassification:

1. instanceof checks first (most reliable, no false positives)
2. error.code checks for Node.js errno codes
3. Message patterns from most specific to least specific (e.g., 'not found' is after 'json' to avoid catching JSON validation errors)
4. Error name fallback for cancellation variants
5. 'unknown' as final fallback

Testing

Added 7 new test cases covering all new error types with real-world error messages from the codebase. All 15 tests pass.