ggml-opencl: drop the FA null-mask assert — it aborts the bidirectional encoder

The GGML_ASSERT(mask != NULL || n_q == 1) added in the previous commit is
wrong: this backend deliberately treats a null mask as bidirectional
(is_causal = 0), and the bidirectional encoder path legitimately dispatches
flash-attn with a null mask and n_q > 1 (the SigLIP/Qwen3-VL vision tower,
n_q = n_kv = 247). The assert therefore fires on a valid call — confirmed by an
on-device SIGABRT during the QVAC-21297 S25 smoke run (ggml_abort in
ggml_cl_compute_forward, right after "flash_attn = enabled").

A null mask + n_q > 1 is indistinguishable from a caller that mistakenly
omitted a causal mask, so there is no safe runtime assertion for this contract;
the existing "null mask ⇒ bidirectional" INVARIANT comment is the correct
documentation. Keep the barrier/divergence fix and the ggml_cl_upscale
zero-source-dim guard.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: iancris

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -9883,7 +9883,6 @@ static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, co
     // that needs causal masking MUST supply an explicit causal mask; relying on
     // shape inference here will silently produce bidirectional (wrong) output.
     const int is_causal = 0;
-    GGML_ASSERT((mask != NULL || n_q == 1) && "FA null mask + n_q>1 requires an explicit causal mask");
 
     const int n_head_log2_val = n_head > 0 ? 1u << (int)floorf(log2f((float)n_head)) : 0;
     const float n_head_log2_f = n_head_log2_val > 0 ? (float)n_head_log2_val : 1.0f;