CANN: update docker images to 8.5.0 and improve CANN.md

[KokerZhou]

This PR includes small updates to CANN-related components:

• update docker images from rc versions to stable 8.5.0

  • this helps ensure a more reliable environment and enables further optimization using newer ops available in stable releases

• update CANN.md to clarify device naming

  • the original document refers to "Atlas 300T A2"
  • however, only "Atlas 300I A2" appears to be officially available
  • the documentation has been adjusted accordingly

These changes are limited to CANN environment setup and documentation.
No functional behavior outside of CANN is affected.

[hipudding]

Thank you very much for your submission. This is something we've been wanting to do recently. llama.cpp only verifies the latest CAN version, and timely updates to the CAN version will result in better compatibility.

[hipudding]

BTW. bf16 is supported by #20152, could you please also add this data type in CANN.md? Thanks.

[KokerZhou]

Clarified BF16 support for 910B. Leave it empty for 310P.
On 310P, BF16/Q8_0/Q4_0 run but may be slower than FP16, as they do not leverage NPU acceleration (likely CPU-backed), as reported in #18797.
BTW, it seems that aclnnWeightQuantBatchMatmulV2 can now work on 310P under certain conditions. The issue mentioned there can be handled carefully later.

[hipudding]

Clarified BF16 support for 910B. Leave it empty for 310P. On 310P, BF16/Q8_0/Q4_0 run but may be slower than FP16, as they do not leverage NPU acceleration (likely CPU-backed), as reported in #18797. BTW, it seems that aclnnWeightQuantBatchMatmulV2 can now work on 310P under certain conditions. The issue mentioned there can be handled carefully later.

Thanks for your feedback. Q8 and Q4 are slow even it run on NPU, not recommend to use 310p with these dtypes.