kernels.cu is about 40% dead code and ffi.rs no longer mirrors it

[CanReader]

The module doc on src/cuda/ffi.rs says it holds "every raw declaration of a
symbol in cuda/kernels.cu, in one place", and that adding a kernel means adding
its signature there and a wrapper in kernels.rs. That's broken in both
directions now.

kernels.cu exports 82 symbols, ffi.rs declares 55. Exported and unreachable:

fastnn_cuda_conv2d_forward / _backward_data / _backward_weight
fastnn_cuda_max_pool2d / _backward, _avg_pool2d, _adaptive_avg_pool2d
fastnn_cuda_batch_norm_forward / _backward
fastnn_cuda_rms_norm_forward
fastnn_cuda_dropout_forward / _backward
fastnn_cuda_cross_entropy_loss, _mse_loss, _binary_cross_entropy
fastnn_cuda_scaled_dot_product_attention
fastnn_cuda_sgd_step, _adam_step
fastnn_cuda_copy, _arange, _where, _gather
fastnn_cuda_leaky_relu_backward

None of them are covered by cuda_parity. One is known broken
(fastnn_cuda_argmax, #9), and fastnn_cuda_conv2d_forward does a D2H
cudaMemcpy per output channel per batch item just to read one bias scalar
(kernels.cu:951). There are also two overloads each of kernel_im2col and
kernel_col2im, and the older pair silently ignores dilation.

A few of these are worth wiring up rather than deleting.
fastnn_cuda_leaky_relu_backward works fine but LeakyReluBackward
(src/autograd/ops/activation.rs:111) still takes the host round trip because the
symbol isn't in ffi.rs. Same for the pooling kernels (#23) and dropout (#20).

Delete what isn't reachable, wire up what should be, and add a CI check that the
two symbol sets match.