metal : split graphs into 4 command buffers on iOS

[jasontitus]

Overview

On iOS, a Metal command buffer that runs for about 5 s of GPU time while another GPU client (such as the display compositor) is waiting gets discarded, and the graph fails with Discarded (victim of GPU error/recovery) (00000005:kIOGPUCommandBufferCallbackErrorInnocentVictim) (llama_decode returns -3). The Metal backend encodes a graph as a small first command buffer plus n_cb = 1 more holding ~90% of the nodes, so a long prefill ubatch on a phone is one multi-second command buffer.

This PR makes the default n_cb 4 on iPhone-class OSes (iOS, iPadOS, visionOS, tvOS; not Mac Catalyst) and keeps 1 elsewhere; GGML_METAL_N_CB=1..8 overrides it. Output is unchanged: splitting only moves command-buffer boundaries.

It also fixes two latent issues that a default above 1 exposes:

• with an abort callback set, graph_compute commits only command buffers 0 and 1 (later ones only when capturing). The main thread's buffer is index n_cb, so n_cb > 1 hung in synchronize. n_cb is now clamped to 1 while an abort callback is set.
• ggml_metal_free released GGML_METAL_MAX_COMMAND_BUFFERS of the + 1 command buffers.

Additional information

iPhone 17 Pro Max (A19 Pro), Ternary Bonsai 2 27B PTQ1_0, 512-token prefill, every run starting at nominal temperature:

Command buffers	Runs	Failed	Longest command buffer
default (1)	50	10	5-8 s (each failure: discarded after 5.0 s)
4	10	0	1.6-2.3 s
1, 256-token ubatches	10	0	3.0-4.3 s

pp512 with 4: 71.7 vs 71.4 tok/s. The decode cost of 4 on the phone is not measured yet; macOS is unchanged (the default stays 1 there).

M5 Max (macOS): full-model logits bitwise identical with GGML_METAL_N_CB 1, 4 and 8 (700 tokens); test-backend-ops MUL_MAT and FLASH_ATTN_EXT pass with 1 and 4; an abort-callback reproducer completes with 1, 4 and 8 (it hung with 4 before the clamp). iOS device build checked.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Claude (Anthropic) helped with the implementation, testing, measurements and this description; I reviewed the change and own it.