cpu: rv64: postops: add rvv postops eltwise op support

[zhangjian29]

Description

This PR adds eltwise operation support for rvv_postops.hpp using a primitive-based approach according to previous proposal #4029.

Motivation

Following the successful integration of primitive-based binary post-ops into rvv_postops.hpp, our next step is to introduce eltwise type post-ops using the same primitive-based paradigm. This is crucial for establishing a general post_ops implementation for the RISC-V 64-bit architecture at this early stage.

Key Changes

• Provided eltwise type of post-ops by integrating with rvv_eltwise in init and execute methods.
• Kept the original post_ops_ok and apply methods unchanged in a fused way, making other implementation used original version still correct.
• Added comments about the original post_ops_ok and apply methods for other future implementations.
• Currently supported post-ops:
  1. relu without alpha/beta (via original methods)
  2. Any eltwise operations the rvv_eltwise supported: relu with alpha/beta, square, abs, sqrt, linear, clip, hardsigmoid, and hardswish.
  3. Any binary operations: add, div, max, min, mul, sub, ge, gt, le, lt, eq, ne, and the ternary select.
• Demonstrated functionality by adding primitive-style eltwise postop support to rvv_nchw_pooling

Future Plans (on-going)

1. Update post-ops support for rvv_matmul in this primitive-based approach.
2. Update post-ops support for rvv_gemm_convolution in this primitive-based approach.
3. Remove the original post_ops_ok and apply methods safely.
4. Add support for sum operations
5. Extend data type supports.

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?

• Have you formatted the code using clang-format?

• test_postops_eltwise_ctest.log

Performance improvements

• Have you submitted performance data that demonstrates performance improvements?

We test the changes using the RISC-V GNU toolchain(14.2) and verify the functionality under the sg2044 platform with cmd:

./tests/benchdnn/benchdnn --pool --dir=FWD_I --attr-post-ops=relu,square,abs,sqrt,linear:0.5:-1,clip,hardsigmoid,hardswish,add:f32,div:f32,mul:f32,max:f32,min:f32,sub:f32,ge:f32,gt:f32,le:f32,lt:f32,eq:f32,ne:f32,select:f32 --batch=../inputs/pool/shapes_basic

All tests passed with proper dispatching confirmed.

• Before the changes: test_postops_without_eltwise.log

• After the changes: test_postops_with_eltwise.log