fix: use LM Studio max_context_length when model is loaded at default 64K

[lunarnexus]

Summary

• fall back to LM Studio max_context_length when the loaded instance context is only Hermes' 64K minimum/default
• keep preferring the loaded instance value when it is above the minimum
• add a regression test covering the default-64K JIT-load case

Problem

Hermes auto-discovers LM Studio context by checking loaded_instances[].config.context_length from /api/v1/models.
When Hermes JIT-loads a model without an explicit context override, LM Studio can load it at the default 64K. Hermes then reports 64K as the model context even when the model's advertised max_context_length is much higher.

Fix

If the loaded instance context is only Hermes' default/minimum load size, fall back to LM Studio's advertised max_context_length. If the loaded instance is already above that minimum, keep using the loaded runtime value.

Why this PR is separate from #36852

This fixes the LM Studio JIT-loaded autodiscovery path.
PR #36852 fixes a different layer: custom_providers per-model context_length lookup when LM Studio reports model ids as publisher/slug.

Related to #30178.

[lunarnexus]

Reviving this PR because it addresses a different LM Studio failure mode than #36852.

• This PR (fix: use LM Studio max_context_length when model is loaded at default 64K #45037): fixes runtime autodiscovery for JIT-loaded LM Studio models that come up at the default 64K, causing Hermes to underreport the model's real context window.
• PR fix(custom_providers): resolve per-model context_length for publisher/slug model ids #36852: fixes custom_providers model-id matching for publisher/slug ids.

So this one is for the JIT/default-64K autodiscovery path, not the config-override path.

[lunarnexus]

Closing this in favor of the consolidated PR #52188, which includes this autodiscovery/reporting fix plus the related LM Studio runtime preload fix from #48884 in one reviewable patch.