fix: retry failed manager initialization

[StellaHuang95]

Summary

Make all six built-in Python environment manager initialize() methods retryable after a failed initialization attempt. A successful initialization stays shared and runs once, and each manager keeps its existing throw-vs-swallow behavior.

Problem

Every manager memoizes initialization in a _initialized deferred and, in finally, resolves it and leaves it set even when discovery threw. The top-of-method guard if (this._initialized) return this._initialized.promise; then returns that settled promise forever, so a single transient failure (e.g. a tool momentarily unavailable during the first call) permanently poisons discovery for the rest of the session with no way to retry.

Affected managers: venv, system, conda, pipenv, poetry, pyenv.

Reproduction: call initialize() while discovery fails (venv/system reject; conda/pipenv/poetry/pyenv swallow and log), then fix the environment and call initialize() again. No rediscovery happens and the manager stays empty for the session.

Root cause

finally { this._initialized.resolve(); } runs on both success and failure and never clears the guard, so the guard short-circuits every later call.

Fix

Mirror the existing reset-on-failure pattern in fastPath.ts. In each initialize():

• Capture the deferred locally: const initialized = createDeferred<void>(); this._initialized = initialized;.
• On an actual thrown exception, clear the guard so a later call retries, but only if this run still owns it: if (this._initialized === initialized) { this._initialized = undefined; }.
• Always settle the captured deferred in finally (initialized.resolve()) so concurrent waiters unblock.

State transitions:

• success: undefined -> deferred -> resolved (stays set; shared, one-time).
• failure: undefined -> deferred -> resolved with the guard cleared (retryable).
• non-throwing tool_not_found / manager-absent: treated as completed init (no wasteful rediscovery).

Preserved behavior:

• throw-style (venv, system) still rethrow; swallow-style (conda, pipenv, poetry, pyenv) still log, emit telemetry, and never throw to callers.
• the ownership check (=== initialized) stops a late failing run from clearing a guard that a concurrent clearCache() + reinit installed.

Tests

New focused initialize() suites for venv, system, pyenv, pipenv, and poetry, plus an extended conda suite. They cover:

• a failed run clears state so a later call retries and succeeds;
• a successful init is not re-run;
• concurrent callers share one run and all settle (throw-style leader rejects while waiters resolve; swallow-style all resolve);
• non-throwing tool_not_found is treated as completed and is not retried;
• a late failing run does not clobber a newer deferred installed by clearCache() + reinit.

[StellaHuang95]

🔒 Automated review in progress — @StellaHuang95 is auto-reviewing this PR.