fix: preserve active model tuning across context-window reloads (#6200)

[atomantic]

Summary

Fixes ensureContextWindow in server/services/ollamaManager.js so that a context-window reload composes on top of the currently applied launch environment (appliedLaunchEnvValues) instead of replacing it with {}. This ensures active model tuning knobs (such as OLLAMA_FLASH_ATTENTION, OLLAMA_KV_CACHE_TYPE) survive context-window reloads triggered by agent spawns (e.g. claude-ollama requiring 128K context).

Additionally, preserves the tuning's undo bookkeeping (preTuningEnv) across non-tuning reloads and updates the baseline's context window, so that when a tuning sweep later completes and calls clearLaunchEnv(), it cleanly restores the untuned daemon with the updated context window rather than reverting it or stranding the tuning.

Test plan

• Verified that all 72 tests in server/services/ollamaManager.test.js pass with Vitest.
• Added regression tests in server/services/ollamaManager.test.js:
  • preserves an active tuning across ensureContextWindow on a spawned process
  • preserves an active tuning and its undo bookkeeping across a service context-window reload
  • preserves an active tuning when ensureContextWindow starts an offline daemon
• Ran related test suites: server/lib/ollamaContext.test.js, server/lib/localModelTuning.test.js, server/services/ollamaAgentContext.test.js, server/services/localModelAssessments.test.js (all passing).

Closes #6200

[atomantic]

Required code review was not completed before publication. This PR is intentionally left open and will not be merged until the required review completes.