fix: use BillingModel and upstream fallback for GPT billing in OpenAI gateway

[octo-patch]

Fixes #1430

Problem

When using Claude Code (CC) with a GPT account through the /v1/messages endpoint, billing could show zero or incorrect amounts:

1. Billing zero: When the account's model mapping uses a non-standard model name (e.g. gpt without a version number), RecordUsage fails to find pricing and silently sets cost to 0. This is because normalizeCodexModel maps gpt → gpt-5.1 (for the upstream request) but the billing logic uses the original un-normalized gpt name which has no pricing entry.

2. Wrong billing model on Messages path: The Anthropic Messages compat path (used by CC) correctly sets result.BillingModel to the GPT billing model (e.g. gpt-5.1-codex), but RecordUsage ignored this field and used result.Model (the Claude model name, e.g. claude-opus-4-6) for pricing instead, resulting in billing at Claude rates rather than GPT rates.

Solution

Two targeted changes to OpenAIGatewayService.RecordUsage:

1. Prefer result.BillingModel when it is explicitly set, falling back to the existing forwardResultBillingModel logic. This ensures the Messages compat path bills at the correct GPT rate.

2. Upstream model fallback: When the resolved billing model has no configured pricing, try result.UpstreamModel before giving up. The upstream model is always normalised to a known model (e.g. gpt-5.1), so this prevents billing from silently dropping to zero for non-standard model names.

Testing

• Added TestOpenAIGatewayServiceRecordUsage_UnknownBillingModelFallsBackToUpstream: verifies that when billingModel="gpt" (no pricing) and upstreamModel="gpt-5.1" (has pricing), cost is non-zero.
• Added TestOpenAIGatewayServiceRecordUsage_BillingModelPreferredOverRequestedModel: verifies that when BillingModel="gpt-5.1-codex" and Model="claude-opus-4-6", billing uses GPT pricing.