fix: self-heal stale url-ingestion provider state

[Wallgau]

This PR fixes URL-ingestion failures caused by stale provider/runtime state in Langflow and MCP wiring.

• Scopes ingestion and MCP provider credentials to the selected embedding provider instead of sending all provider vars.
• Removes stale provider headers (for example old OLLAMA_BASE_URL) from MCP server args so non-selected providers are not accidentally used.
• Adds conditional self-heal for URL ingestion: on known stale-state signatures (for example “Failed to connect to Ollama” while provider is not Ollama), reconcile once and retry once.
• Improves retry-failure errors with provider-aware messaging so failures are easier to diagnose.

Issue: #1170

• Users could still hit “Failed to connect to Ollama” even when OpenAI was selected, due to stale runtime/MCP state and mixed provider headers. This PR makes provider selection authoritative and adds safe, targeted recovery.

What Changed

• src/utils/langflow_headers.py
• Provider-scoped credential/header/global-var generation.
• src/services/langflow_mcp_service.py
• Prunes stale provider headers not in the desired selected-provider set.
• src/services/langflow_file_service.py
• Uses selected provider when building ingestion headers.
• Supports dynamic URL-ingest flow-id resolution when available.
• Adds conditional stale-state reconcile + single retry path.
• Adds provider-specific retry error formatting.

Tests
Added TDD coverage with mocked HTTP (CI-safe, no real external calls):

• tests/unit/test_url_ingestion_pipeline_tdd.py
• selected-provider-only ingestion headers
• prefers resolved flow-id over stale configured id
• reconcile-and-retry once on known stale-state errors
• no retry on non-reconcilable errors
• provider-specific error message after retry failure
• integration-style mocked chain for provider selection + self-heal

tests/unit/test_mcp_provider_selection_tdd.py

• provider-scoped MCP global vars
• stale provider header pruning
• server update applies selected-provider vars only

Before user will get error trying to ingest url:

Now urls are ingest properly and reflect in the knowledge table: