fix: add Qwen3.5 model context window tokens

[alievrusik]

Summary

• Add qwen3.5 (262,144) and qwen3.5-397b (262,144) entries to the qwenModels token map

Problem

Qwen3.5 models (e.g. Qwen/Qwen3.5-397B-A17B-FP8) have a 262,144 token native context window, but were falling back to the generic qwen3 entry (40,960 tokens) via findMatchingPattern fuzzy name matching.

This caused the agent graph's pruneMessages (in @librechat/agents) to aggressively drop messages — including the user query, assistant tool_calls, and tool results — when tool output exceeded the undersized token budget (~36K tokens after the 0.9x scaling formula), leading to the model receiving only the system message and producing broken or empty responses.

Symptoms observed

• Small tool outputs worked fine (~54KB / ~13K tokens fit within 36,864 budget)
• Large tool outputs (~144KB / ~36K tokens) caused the entire conversation to be pruned
• Model responded without context, ignoring tool results or producing "No user query found" errors
• Same scenario worked correctly with Anthropic (which has its own token map entry)
• Direct vLLM API calls with the same large payload worked correctly

Test plan

• Verified Qwen/Qwen3.5-397B-A17B-FP8 now resolves to 262,144 tokens instead of 40,960
• Tested with large tool output (~144KB) — all messages preserved, model responds correctly
• Tested with small tool output — still works as before