Splash: allow contexts above 262,144 tokens (YaRN) and pass --ple-memory

[DPalagi]

Environment: oMLX 0.7.3 (macOS app), Splash 1.5.0, macOS 27.2, Apple M5 Max 128 GB.
Model: ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF (IQ1_M, 2 split files) with incoai/Qwen3.8-Flash-Next-DFlash2.

What I'd like

Two options that Splash 1.5.0 already supports but oMLX does not expose when it launches splash serve:

1. Context above 262,144 tokens. engine/splash.py caps the requested context with
   min(options.max_context, SPLASH_MAX_CONTEXT) where SPLASH_MAX_CONTEXT = 262144, and docs/Splash.md says
   "(at most 256K)". Splash itself accepts --max-context up to 1,010,000 and turns on static YaRN past the native
   window (Qwen's model card: "262,144 natively and extensible up to 1,000,000 tokens").
2. --ple-memory. Splash can keep a small RAM row cache of the 27 GB per-layer n-gram (PLE) table of a
   Flash-Next GGUF and stream the rest from the SSD. oMLX never passes it, so with a full budget the table loads
   into memory (--ple-memory auto). The qwen4_ple_ssd_offload model setting only applies to MLX checkpoints.

Today the only way to get either is to run splash serve by hand outside oMLX, which loses the oMLX dashboard,
profiles, idle unload and memory accounting for that model.

Measurements (same GGUF, splash serve started directly)

Setup	Decode, 2 parallel requests	Prefill 128K	Peak memory (all-in)
Through oMLX (PLE table in RAM)	81 tok/s	1,616 tok/s	~61 GB
--ple-memory 2G	89 tok/s	1,795 tok/s	34-40 GB
--ple-memory 2G --max-context 750K (YaRN ×2.93), 32K context per request	59 tok/s	1,704 tok/s	40-46 GB

• Streaming the PLE table cost nothing measurable and saved ~20-25 GB.
• Long-context recall (5 keys planted at 5-95 % depth in a real Kotlin code base): 5/5 at 128K, 250K, 480K and
  730K with --max-context 512K or 750K. 500K prefill: 1,381 tok/s (TTFT 6 min); 730K: TTFT 9 min.
• Short-context regression with YaRN 750K on: our agent coding suite stayed at 9/9 (same as native).

Possible shape

• Raise SPLASH_MAX_CONTEXT to Splash's own limit (1,010,000), or only cap at the native window unless the model's
  max_context_window setting asks for more — the same rule patches/qwen35_yarn.py already uses for MLX
  qwen4_exp models (yarn_factor = window / native).
• A per-model setting for the PLE row cache (e.g. reuse qwen4_ple_ssd_offload for Splash GGUFs and map it to
  --ple-memory <size>), and count only the cache in the memory estimate. That would also change the estimate
  discussed in Splash Split GGUF incorrect memory requirement #4876.

Related: #4677 (leave the context to Splash when the budget cannot hold the limit), #4876 (Splash split GGUF memory
requirement), #1077 (first-class YaRN config, MLX side).

Happy to test a build or share the full benchmark data.