[Conv] Fix filter iterator pointer offset scaling (issue #3504)

[XFDG]

Summary

Fix add_pointer_offset byte scaling in five convolution filter tile access iterators:

• Conv2d fprop analytic
• Conv2d fprop fixed channels
• Conv2d fprop few channels
• Conv3d fprop analytic
• Depthwise Conv2d fprop direct-convolution optimized

Add a lightweight convolution threadblock unit test that constructs each iterator and verifies its observable pointer displacement.

Rationale

These iterators store their base pointer as char const*, while add_pointer_offset documents its argument in units of Element. The byte displacement must therefore be:

pointer_offset * sizeof_bits::value / 8

The previous inverted expression multiplied by 8 and divided by the element bit width. For fp16, advancing by 8 elements moved 4 bytes instead of 16; fp32 was off by a factor of 16.

Validation

Built the same host/device CUDA probe before and after the change using CUDA 13.1 and -arch=sm_100a, then ran it on an NVIDIA B200.

Before:

• all five fp16 iterators moved 4 bytes for an 8-element offset on both host and device
• 10/10 checks failed

After:

• all five iterators move the expected 16 bytes on host and device
• 10/10 checks pass and the probe exits with failures=0
• the new WITHOUT_CUDA unit-test translation unit compiles successfully with the GPU node host compiler
• git diff --check passes

Fixes #3504