Add support for QRWKV6 hybrid models & slight optimization for RWKV6

[MollySophia]

QRWKV6-32B is a new model by Recursal which is a combination of the Qwen2.5 architecture and RWKV6.
It 'converts' a Qwen2.5-32B-Instruct model's QKV attention into RWKV6 linear attention, keeping knowledges in the origin Qwen model while gaining the advantages of linear models (constant vram usage and flops, independent of ctxlen).
More info/model for testing: https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1
Some converted GGUF for testing: https://huggingface.co/mollysama/QRWKV6-32B-Instruct-Preview-GGUF

Changes in this PR:

• Add OP gated linear attention with CPU and CUDA impl, which looks like a simplified version of RWKV6 wkv attention.
• Model conversion and inferencing for QRWKV6-32B
• RWKV6 optimizations: graph simplification; concated lerp weights to reduce cpu overhead during inference (credit to @compilade)

Testing details:

• 32B Q4_0/Q4_K quantized model running on a single 4090 with decent speed:

$ ./build/bin/llama-bench -m ../QRWKV6-32B-Instruct-Preview-v0.1-Q4_0.gguf -ngl 99
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
| model                          |       size |     params | backend    | ngl |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------------: | -------------------: |
| rwkv6qwen2 32B Q4_0            |  19.34 GiB |    34.74 B | CUDA       |  99 |         pp512 |        819.60 ± 1.01 |
| rwkv6qwen2 32B Q4_0            |  19.34 GiB |    34.74 B | CUDA       |  99 |         tg128 |         32.72 ± 0.01 |

build: 5a73dbcb (4397)

• wikitext2 PPLs:

Quant type	PPL
f32	5.6987 +/- 0.03365
q8_0	5.7005 +/- 0.03370
q6_k	5.7126 +/- 0.03376
q5_k_s	5.7339 +/- 0.03393
q4_k_m	5.7921 +/- 0.03428
q4_0	5.8568 +/- 0.03481
q3_k_m	6.0677 +/- 0.03635
q2_k	7.4547 +/- 0.04597

• Performance of QRWKV6-32B difference before/after concating lerp weights together:

(Sry for the image attachment)

before:
$ ./build/bin/llama-bench -m ../QRWKV6-32B-Instruct-Preview-v0.1/QRWKV6-32B-Instruct-Preview-v0.1-F16.gguf -sm none -mg 0
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 8 CUDA devices:
  Device 0: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 1: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 2: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 3: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 4: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 5: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 6: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 7: NVIDIA H800, compute capability 9.0, VMM: yes
| model                          |       size |     params | backend    | ngl |    sm |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ----: | ------------: | -------------------: |
| rwkv6qwen2 32B F16             |  65.26 GiB |    34.74 B | CUDA       |  99 |  none |         pp512 |        697.64 ± 0.59 |
| rwkv6qwen2 32B F16             |  65.26 GiB |    34.74 B | CUDA       |  99 |  none |         tg128 |         21.91 ± 0.00 |

build: b7b45753 (4397)

after:
$ ./build/bin/llama-bench -m ../QRWKV6-32B-Instruct-Preview-v0.1/QRWKV6-32B-Instruct-Preview-v0.1-F16-fused-lerp.gguf -sm none -mg 0
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 8 CUDA devices:
  Device 0: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 1: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 2: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 3: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 4: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 5: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 6: NVIDIA H800, compute capability 9.0, VMM: yes
  Device 7: NVIDIA H800, compute capability 9.0, VMM: yes
| model                          |       size |     params | backend    | ngl |    sm |          test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ----: | ------------: | -------------------: |
| rwkv6qwen2 32B F16             |  65.26 GiB |    34.74 B | CUDA       |  99 |  none |         pp512 |        731.32 ± 1.10 |
| rwkv6qwen2 32B F16             |  65.26 GiB |    34.74 B | CUDA       |  99 |  none |         tg128 |         26.51 ± 0.01 |

build: b7b45753 (4397)

[MollySophia]

Hi! @ggerganov
May I request for a review? :3

[ggerganov]

I haven't tested the models. ggml-ci is passing on my CUDA machine.

[MollySophia]

Self-reviewed again and fixed a merge issue after the refacor that makes RWKV6 fail to run with fused lerp weights.
I think this PR is ready to go.