vulkan: support arbitrary KV dimension in flash attention

[jeffbolznv]

The "Clamp" spec constant is already based on whether KV is a multiple of Bc, so use that to control whether bounds checking is performed. Add bounds checking to the scalar and coopmat1 paths. Coopmat2 didn't need any changes (the K/V tensors are already optionally clamped, nothing else needed to be changed).

For #16148.

[jeffbolznv]

These accelerated vulkan CI jobs are a pleasant surprise. The failures are unrelated, but I'll look into them.

[0cc4m]

Not the fault of this PR, but I'm getting a validation issue with Flash Attention:

VUID-VkComputePipelineCreateInfo-layout-07991(ERROR / SPEC): msgNum: 989008490 - Validation Error: [ VUID-VkComputePipelineCreateInfo-layout-07991 ] Object 0: handle = 0x2cfba2000000001c, type = VK_OBJECT_TYPE_SHADER_MODULE; Object 1: handle = 0x908683000000001d, type = VK_OBJECT_TYPE_PIPELINE_LAYOUT; | MessageID = 0x3af3126a | vkCreateComputePipe
lines(): pCreateInfos[0].stage SPIR-V (VK_SHADER_STAGE_COMPUTE_BIT) uses descriptor [Set 0, Binding 1] with a VkDescriptorSetLayoutBinding::descriptorCount of 1, but requires at least 2 in the SPIR-V.
The Vulkan spec states: If a resource variable is declared in a shader as an array, the corresponding descriptor set in layout must match the descriptor count (https://docs.vulkan.org/spec/latest/chapters/pipelines.html#VUID-VkComputePipelineCreateInfo-layout-07991)
    Objects: 2
        [0] 0x2cfba2000000001c, type: 15, name: NULL
        [1] 0x908683000000001d, type: 17, name: NULL