[Perf] Implement lazy loading for AI provider SDKs

[randomm]

Problem

All 21 AI provider SDKs are eagerly imported at startup, adding 20-25 MB to the bundle. Most users only use 1-2 providers.

Current State

• packages/opencode/src/provider/provider.ts (lines 17-39)
• Static imports for all 21 providers
• BUNDLED_PROVIDERS map references eager imports

Solution

Convert to lazy loading using dynamic imports:

// Current (eager)
import { createAmazonBedrock } from "@ai-sdk/amazon-bedrock"
const BUNDLED_PROVIDERS = {
 "@ai-sdk/amazon-bedrock": createAmazonBedrock,
}

// Proposed (lazy) 
const BUNDLED_PROVIDERS = {
 "@ai-sdk/amazon-bedrock": () => import("@ai-sdk/amazon-bedrock").then(m => m.createAmazonBedrock),
}

Why It's Feasible

• ✅ getSDK() is already async
• ✅ Fallback pattern exists for custom providers
• ✅ State caching already implemented
• ✅ Zero architecture changes needed

Expected Impact

• Cold start: 20-25 MB → ~500 KB
• Per provider: ~1.5-2.5 MB (on demand)
• Typical user (2 providers): ~4 MB total

Trade-off

• First provider access: +50-100ms latency (one-time)
• Subsequent: instant (cached)

Acceptance Criteria

• All 21 providers converted to dynamic imports
• No startup regression
• Provider caching still works
• Tests pass