feat: add MaxAttachments to Model struct and Gemini/Vertex configs (#258)

[mvanhorn]

Closes #258.

What changed

• pkg/catwalk/provider.go — added MaxAttachments int64 to the Model struct, with json tag max_attachments,omitempty. Documented inline that zero means "no provider-enforced cap" and a positive value is a hard cap the consumer must respect.
• internal/providers/configs/gemini.json — all 6 models get max_attachments: 10.
• internal/providers/configs/vertexai.json — all 11 models get max_attachments: 10.

The proposal in #258 lists Gemini and Vertex AI as the only providers with documented per-request image caps; Anthropic and OpenAI gate by context window, so their configs stay at the zero default.

Why this matters

Consumers like Crush hard-code map[InferenceProvider]int{InferenceProviderGemini: 10, ...} in their own source today. That breaks two things:

1. Granularity: the lookup table is per-provider, so it can't represent a hypothetical future Gemini model with a different limit.
2. Release coupling: every time a provider changes a limit, every consumer ships a release.

Carrying the limit on Model itself moves it to the same boundary that already carries context_window, default_max_tokens, cost_per_1m_in, etc. The omitempty tag keeps existing consumers that don't know about the field unaffected.

Verification

• go build ./... — clean
• go test ./... — internal/providers test passes (TestValidDefaultModels)
• Both updated JSON configs parse cleanly and report max_attachments=10 on every model