VLM: fix EmbeddedInput shape handling, error recovery, and parallel loading

[stikves]

Summary

• EmbeddedInput.tokenCount: correctly handle 2D [seq_len, hidden_dim] tensors (was returning hidden_dim instead of seq_len). Now a stored property computed once at init.
• EmbeddedInput.seqLen(of:): centralized helper for seq_len extraction — scatterMerge and other call sites use one canonical path instead of inline shape checks.
• scatterMerge: replace precondition(float16) with a thrown error for bfloat16 model compatibility.
• LLMRunnerMain: add comment explaining sequential PreparedModel.prepare (parallel caused runtime errors).
• Removed premature --max-tiles flag (tiling model not implemented yet).

Test plan

• Verify tokenCount returns correct seq_len for both 2D and 3D embeddings
• Confirm bfloat16 model gets a catchable error instead of a crash
• VLM inference still works end-to-end

[carinapeng]

This looks good to me overall but before approving I think a few things should be fixed, one is on how we read maxTiles, and the other is consistency - seems correct for the 2d case but scatter merge still issues 3D: [1, seq, hidden]

And let's do the testing on the Qwen3 VL as well since that is released :)

[carinapeng]

Should we rebase this PR?

Also an alternative to the max tiling is to restore the missing property here in LLMRunnerMain.swift, we'd get a build failure right now on ToT main which breaks running VLMs

      183 +        name: .customLong("max-tiles"),
      184          help: "Maximum tiles for image splitting (overrides model config). 1 = single crop, no tiling.")
      185 +    var maxTiles: Int?

[carinapeng]

LGTM, let's merge after verifying commits!