Feat/dsv4 perf model

[devalshahamd]

This pull request adds support for DeepSeek-V4 sparse paged decode and the mHC (manifold-constrained Hyper-Connection) operator family to the performance modeling codebase. It introduces new perf model classes for these operators, updates the pseudo-op mappings to recognize them, and extends category-only mappings.

DeepSeek-V4 Sparse Attention Support:

• Added _V4PagedDecodeBase and three subclasses (pseudo_v4_paged_decode_swa, pseudo_v4_paged_decode_csa, pseudo_v4_paged_decode_hca) to model different DeepSeek-V4 paged decode modes, plus a perf model for aiter::pa_sparse_prefill_opus_fwd. These provide detailed parameter extraction, FLOP, and byte estimation for DeepSeek-V4 attention ops.
• Registered these new classes in the pseudo-op mapping so they are recognized and used for the corresponding synthetic op names.

mHC (manifold-constrained Hyper-Connection) Operator Family:

• Implemented a new _MHCBase class and five subclasses (mhc_post, mhc_pre_gemm_sqrsum, mhc_pre_big_fuse_rmsnorm, mhc_fused_post_pre_gemm_sqrsum, topk_softplus) to model the main mHC and MoE auxiliary ops, including parameter extraction, FLOP, and byte estimation.
• Registered these mHC and MoE auxiliary perf models in the pseudo-op mapping for their corresponding op names.

Synthetic Op Category Classification:

• Extended category-only mappings to classify small ops (including indexer, FusedRoPE, and custom collective variants) by name, even when shape parameters are not present.