fix(chat): reserve at the long-context step for huge prompts

[VickyXAI]

estimateChatCost reserved at base rates for every prompt size. Past a model's pricing.long_context threshold, the gateway reprices the whole request, so a huge prompt under-reserved. Examples: gpt-6-astra over 272K reserved $10/$50 against a $20/$75 quote, and Grok from 200K reserved half the quote.

• LONG_CONTEXT_PRICE_PER_MTOKEN copies the live steps from GET /v1/models (both chains agree; the Qwen steps are Base-only). That covers GPT-6, GPT-5.4/5.5/5.6 (>272K, 2x in / 1.5x out), Gemini 3.1/2.5 Pro (>200K, 2x/1.5x), Grok 4.3–4.7 and build-0.1 (>=200K, 2x/2x), and Qwen 3.7 plus/flash (inclusive; flash steps twice).
• chatRateAt / tierWorstPriceAt pick the highest step reached, using the gateway's own inclusive/exclusive rule. They are pure in prompt tokens and never lower the base rate. They are used by estimateChatCost for both an explicit model and a mode, and by accountLedgerUsd.
• Tests cover the exact boundaries: 272,000 stays base and 272,001 steps for the exclusive models; 199,999 stays base and 200,000 steps for inclusive Grok. They also cover the multi-step ladder, tier modes, bare ids, prototype keys and the ledger.
• verify:prices now fails when a live step would settle above the reserve. A control run with the astra step removed and grok made exclusive flagged both models on both chains.

Refs BlockRunAI/blockrun#781