Fix and accelerate LLM tabular background runs

[Paul Lizer (paullizer)]

Summary - Fixes streaming direct source-backed tabular queue calls by passing required model and settings arguments with explicit keywords. - Automatically retries LLM row-count/schema/source-token validation failures, then preserves manual Continue from durable checkpoints after automatic retries are exhausted. - Keeps every row-level answer LLM-generated while increasing independent model-call concurrency through adaptive 4/16/64/128 tiers; explicit administrator concurrency still overrides the tier. - Sizes source batches from selected-model metadata or optional input/output token-limit fields in the local model capability catalog, with bounded fallback budgets, 50% input and 60% output targets, and a 180k soft input cap for contexts over 500k tokens. - Calculates ETA from recent wall-clock rows/minute across parallel windows instead of summing concurrent request durations as serial work. - Persists progress once per completed parallel window and exposes throughput, configured concurrency, and effective concurrency on the chat status card. - Bumps the application to 0.250.136 and updates feature/test documentation. ## Production evidence - The 30,000-row source staged as 909 batches of about 33 rows under the prior 60k-character budget. - Runtime used concurrency 3 and reported about 35 hours remaining because concurrent request durations were serialized in ETA math. - At the observed roughly 155 seconds per 33-row model call, 100-128 effective asynchronous calls project the same LLM-only workload into roughly the 18-25 minute range before validation retries. Actual time remains model- and output-volume-dependent. ## Validation - python -m py_compile application/single_app/config.py application/single_app/functions_tabular_generated_exports.py functional_tests/test_tabular_row_orchestration_scale.py ui_tests/test_chat_background_generated_export_status.py - python functional_tests/test_tabular_row_orchestration_scale.py - node --check application/single_app/static/js/chat/chat-messages.js - python -m json.tool application/single_app/static/json/model_capabilities.json - python -m pytest ui_tests/test_chat_background_generated_export_status.py -q (5 skipped locally because authenticated UI target variables are not configured) - VS Code diagnostics: clean for all changed code/test files - git -c core.whitespace=blank-at-eol,blank-at-eof,space-before-tab,cr-at-eol diff --check

Refs #1031

[Paul Lizer (paullizer)]

Added retry capability for failed tabular generated exports:

• Model-output validation failures now auto-requeue up to abular_generated_output_model_validation_auto_retries (default 3).
• After the auto budget is exhausted, the run is marked failed but remains retryable so the existing manual Continue action can resume from checkpoints.
• Scheduler auto-claim skips exhausted failures; manual Continue clears the exhausted state and queues another attempt.
• Added regression coverage for auto retries followed by manual-only continuation.

Validation rerun:

• .\.venv\Scripts\python.exe -m py_compile application\single_app\functions_tabular_generated_exports.py application\single_app\functions_settings.py application\single_app\config.py functional_tests\test_tabular_row_orchestration_scale.py
• .\.venv\Scripts\python.exe functional_tests\test_tabular_row_orchestration_scale.py