Report prefill progress on the VLM path

[solderzzc]

Summary

On --vision models, n_past and fraction in the prefill progress heartbeat stayed at 0 for the whole prefill. Only the LLM prepare calls activePrefillProgressHook. The VLM prepare / prepareContinuation report through PrefillParameters.progress, which SwiftLM never set.

• Set prefill.progress on the chat and text completion parameters to forward into activePrefillProgressHook.
• PrefillState.update keeps the highest nPast, because updates arrive as detached Tasks and can land out of order.

Test plan

• Qwen3.8-27B-4bit --vision, 8,413-token prompt: before, n_past stayed 0 for 26 s; after, it rose 495 → 7,920 (fraction 0.06 → 0.94), and the answer was correct
• swift test --filter SwiftLMTests: 185 tests, 0 failures

🤖 Generated with Claude Code

[solderzzc]

Cross-review (M6 session): no blocking issues. OK to merge once CI is green.

Reviewed the full diff (Server.swift, +17/−3) against main 804616e:

• Both GenerateParameters constructions (chat and text completions) now forward prefill.progress, and each ends in a let params snapshot, so there's no captured-var warning. A clean rebuild of 499c36c on Xcode 27 / Swift 6.4 shows none from this PR.
• On the LLM path, LLMModel.prepare still calls activePrefillProgressHook directly, and prefill.progress may forward the same (upperBound, total) again. With PrefillState.update now taking max(nPast, …), the duplicate and any out-of-order detached updates are harmless.
• Not introduced here, noted for later: activePrefillProgressHook is a process-global, so with --parallel > 1 one request's prefill could drive another's heartbeat. That's pre-existing, and this PR doesn't make it worse.

Verified on a Mac mini M6 (32 GB): Qwen3.8-27B-4bit (VLM path), --ctx-size 16384, 4.9K-token prompt with X-SwiftLM-Prefill-Progress: true. The heartbeat's n_past goes 492 → 984 → 1476 → 1968 → 2952 → 3444 → 3936 → 4428 and fraction 0.1 → 0.9, every 2 s, monotonic, and the answer is correct. Before this PR, n_past stayed at 0 on the VLM path.