fix(gateway): read context_length from custom_providers in session info header

[JanCong]

Problem

_format_session_info() in gateway/run.py only reads context_length from the top-level model.context_length in config.yaml. If a user configures context_length under custom_providers → specific provider → models.<model>.context_length, this is completely ignored, and the session header always shows "128K tokens (default — set model.context_length in config to override)".

Fix

Added a scan of custom_providers entries when the top-level model.context_length is not set. The resolution order is now:

1. Top-level model.context_length (existing, highest priority)
2. custom_providers → matching provider → models.<model>.context_length
3. Fall through to the existing probe/default behavior

The fix is a ~35 line addition that iterates through custom_providers looking for context_length matching the current model. Also checks both cp[context_length] (provider-level) and cp[models][model][context_length] (per-model).

Testing

• All 19 gateway-related tests pass
• 250 existing run_agent tests pass (3 pre-existing unrelated failures)

[alt-glitch]

Likely duplicate of #15668, #5096, #8240 — all fix _format_session_info() to read context_length from custom_providers. Note that #15844 (merged) already fixed the runtime resolution path; this addresses the remaining display-only gap in the session info header.

[teknium1]

Salvaged via #19708 onto current main. A custom_provs hoist added on main since your PR opened was preserved alongside your data init, so both your session-info fix and the downstream custom-provider usage sites continue to work. Thanks @JanCong!