[Feature] Add statistical input token admission for AI functions

Feature request

Is your feature request related to a problem? Please describe.

AI SQL functions can invoke an external model once per input row, so an unexpectedly large query can consume significant input tokens. StarRocks currently has no pre-execution budget for this estimated work.

Describe the solution you'd like

Add an optional statistical input-token admission check using the existing physical AIProject plan and optimizer statistics:

• Show the plan's estimate, or an explicit unknown result, in EXPLAIN VERBOSE/COSTS.
• Keep admission disabled by default; when enabled, reject unknown or over-budget estimates before deployment.
• Preserve the existing AI Provider and SYSTEM execution routes and query/transaction lifecycle.
• Document supported estimates and their limitations: this is not exact tokenization, a billing quota, or runtime token accounting.

Describe alternatives you've considered

Query queue CPU/memory accounting and runtime usage reporting serve different purposes: neither prevents a plan from starting based on estimated external-model input tokens. This feature should reuse existing statistics and deployment boundaries without changing those mechanisms.

Additional context

Related AI function integration: #78807.