RoPE: fix back, CUDA support for back + noncont.

[JohannesGaessler]

This PR fixes the backward pass for RoPE. On master test-backend-ops grad is failing on a related assert. The backwards pass can be constructed relatively simply by just creating a tensor for the forward pass and then changing the op from GGML_ROPE to GGML_ROPE_BACK. One could maybe set ggml_tensor.op_params instead of the op but I don't think that would reduce the overall complexity.

This PR also adds CUDA support for the RoPE backwards pass and for non-contiguous inputs. The latter is needed for the backwards pass of the KV cache. I also added __restrict__ and const where applicable and simplified the templating a bit.

Implicitly test-backend-ops is already testing GGML_OP_ROPE_BACK via gradients, I also added an explicit test.

[ggerganov]

Now that non-contiguous rope is supported, the extra ggml_cont in build_minicpm3 and build_deepseek2 could likely be removed:

llama.cpp/src/llama.cpp

Lines 6499 to 6514 in 39509fb

	q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
	q_pe = ggml_rope_ext(
	ctx0, q_pe, inp_pos, nullptr,
	n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
	ext_factor, attn_factor_scaled, beta_fast, beta_slow
	);
	cb(q_pe, "q_pe", il);
	// shared RoPE key
	k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
	k_pe = ggml_rope_ext(
	ctx0, k_pe, inp_pos, nullptr,
	n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
	ext_factor, attn_factor_scaled, beta_fast, beta_slow
	);
	cb(k_pe, "k_pe", il);

But it would be good to test these models just in case.
If you don't have the models handy, you can simply update the TODO comments with one that asks to perform the test, so that somebody else can do it in the future.

[ggerganov]

The build_minicpm3 has the same comments that should be updated