The prebuilds reconcile loop can currently spam workspace creation
attempts that are always going to fail, such as when required dynamic
parameters are missing or unresolved.

This change introduces an in-memory per-preset backoff mechanism that
tracks consecutive creation failures and delays subsequent creation
attempts using linear backoff:

- First failure: backoff for 1x interval (default 1 minute)
- Second consecutive failure: backoff for 2x interval (2 minutes)
- Third consecutive failure: backoff for 3x interval (3 minutes)
- And so on...

When a creation succeeds, the failure tracking is cleared and any
subsequent failure starts backoff from 1x interval again.

This complements the existing database-based backoff system by
preventing immediate retry spam when creation fails quickly (e.g.,
due to missing parameters), while still allowing periodic retries
and recovery when issues are fixed.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Encapsulates the failure tracking logic into a dedicated
presetFailuresTracker struct with its own methods, improving code
organization and separation of concerns.

Changes:
- Created presetFailuresTracker struct with failures map and mutex
- Moved RecordFailure, RecordSuccess, and ShouldBackoff methods to
  the tracker
- Updated StoreReconciler to hold a failureTracker instance
- Updated all call sites to use the new tracker methods

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>