Flash Attention disabled mode support

[ravi9]

Ensure models run correctly when Flash Attention is unavailable by providing alternative implementations.

[wine99]

Paste comments from #118:

Verified -fa off with llama-cli, llama-bench, and llama-perplexity on CPU.

Two cases are still failing:

• -fa off with STATEFUL
• NPU

STATEFUL and NPU support will be addressed in separate PRs.

When -fa off, v_trans=true. In this case, the layout of cache V differs from cache K. Although the tensor shape of cache V is the same as cache K, i.e. [n_embd, kv_size, n_seq, 1], the data is actually stored and accessed in a transposed manner. The effective layout is [kv_size, n_embd, n_seq, 1].

This causes two issues:

1. The SET_ROWS op that updates cache V first flattens the cache and writes the current V using a scattered, element-wise update pattern, instead of row-wise updates as in cache K. Code snippets:

-  13: [   512,     2,     1,     1] RESHAPE              Vcur-0 (reshaped)                            [ 4,  2048,  4096,  4096]       OPENVINO0
       [    64,     8,     2,     1]            0: RESHAPE     Vcur-0                                      [ 4,   256,  2048,  4096]       OPENVINO0
-  14: [     1,  1024,     1,     1] RESHAPE              Vcur-0 (reshaped) (reshaped)                 [ 4,     4,  4096,  4096]       OPENVINO0
       [   512,     2,     1,     1]            0: RESHAPE     Vcur-0 (reshaped)                           [ 4,  2048,  4096,  4096]       OPENVINO0
-  15: [     1, 524288,     1,     1] RESHAPE              cache_v_l0 (reshaped)                        [ 2,     2, 1048576, 1048576]       OPENVINO0
       [   512,  1024,     1,     1]            0: NONE        cache_v_l0                                  [ 2,  1024, 1048576, 1048576]       OPENVINO0
-  16: [     1, 524288,     1,     1] SET_ROWS             cache_v_l0 (reshaped) (view)                 [ 2,     2, 1048576, 1048576]       OPENVINO0
       [     1,  1024,     1,     1]            0: RESHAPE     Vcur-0 (reshaped) (reshaped)                [ 4,     4,  4096,  4096]       OPENVINO0
       [  1024,     1,     1,     1]            1: NONE        leaf_11                                     [ 8,  8192,  8192,  8192]  OPENVINO0_HOST
       [     1, 524288,     1,     1]            2: RESHAPE     cache_v_l0 (reshaped)                       [ 2,     2, 1048576, 1048576]       OPENVINO0

2. The VIEW–PERMUTE chain applied to cache V before attention differs from that of cache K. Code snippets:

-  53: [   256,     8,    64,     1] VIEW                 cache_v_l1 (view)                            [ 2, 131072,  2048, 1048576]       OPENVINO0
       [   512,  1024,     1,     1]            0: NONE        cache_v_l1                                  [ 2,  1024, 1048576, 1048576]       OPENVINO0
-  54: [   256,    64,     8,     1] PERMUTE              cache_v_l1 (view) (permuted)                 [ 2,  2048, 131072, 1048576]       OPENVINO0
       [   256,     8,    64,     1]            0: VIEW        cache_v_l1 (view)                           [ 2, 131072,  2048, 1048576]       OPENVINO0
-  55: [    64,     8,   256,     1] VIEW                 cache_k_l1 (view)                            [ 2,   128,  1024, 1048576]       OPENVINO0
       [   512,  1024,     1,     1]            0: NONE        cache_k_l1                                  [ 2,  1024, 1048576, 1048576]       OPENVINO0
-  56: [    64,   256,     8,     1] PERMUTE              cache_k_l1 (view) (permuted)                 [ 2,  1024,   128, 1048576]       OPENVINO0
       [    64,     8,   256,     1]            0: VIEW        cache_k_l1 (view)                           [ 2,   128,  1024, 1048576]       OPENVINO0

STATEFUL is currently not working due to shape and layout mismatches.

NPU is not working because it pads both the current K/V tensors and the SET_ROWS index input. This approach only works with row-wise updates.