refactor(ops): remove InfiniLM-suffixed InfiniOps calls

[voltjia]

What

• Remove the InfiniCore adapters that call the InfiniOps KvCachingInfinilm, PagedAttentionPrefillInfinilm, RandomSampleInfinilm, and RotaryEmbeddingInfinilm APIs.
• Replace the PagedAttentionInfinilm decode fallback with the existing InfiniOP paged-attention provider while retaining canonical FlashAttnWithKvcache dispatch for supported layouts.
• Simplify linked-operator selection so it only tracks the canonical FlashAttnWithKvcache and FlashAttnVarlenFunc APIs.
• Remove every InfiniLM-suffixed InfiniOps reference from InfiniCore source and build configuration.

Migration

InfiniCore route	Removed InfiniOps call	Replacement
kv_caching	KvCachingInfinilm	Existing InfiniOP KV-caching provider
paged_attention_prefill	PagedAttentionPrefillInfinilm	Existing InfiniOP paged-attention-prefill provider
random_sample	RandomSampleInfinilm	Existing InfiniOP random-sampling provider
rope	RotaryEmbeddingInfinilm	Existing InfiniOP rotary-embedding provider
Unsupported paged_attention decode layouts	PagedAttentionInfinilm	Existing paged_attention_impl::infiniop::{plan, run, cleanup} path
Supported paged_attention decode layouts	N/A	Canonical InfiniOps FlashAttnWithKvcache path retained
Paged mha_varlen layouts	N/A	Canonical InfiniOps FlashAttnVarlenFunc path retained

Scope

No public InfiniCore Python or C++ API changes are introduced, and the InfiniOps gitlink is unchanged. The existing InfiniOP implementations become the fallback providers; this PR does not add another operator implementation.

This PR is stacked on #1481 (refactor/migrate-paged-attention-infiniops) so its diff contains only the suffix-removal work. It can be retargeted to main after #1481 merges.

Screenshots: N/A (backend adapter migration only).

Validation

Run on ssh nvidia in accelerator-dev/nvidia:latest on NVIDIA A100 GPUs:

• git diff --check passed.
• clang-format 16.0.6 passed for the modified C++ source.
• The canonical linked InfiniOps build resolved and built only FlashAttnWithKvcache and FlashAttnVarlenFunc; full InfiniCore shared-library builds passed.
• Dynamic-symbol scans of libinfinicore_cpp_api.so, _infinicore, and libinfiniops.so found no Infinilm symbols.
• Source scan of src/infinicore and xmake.lua found no InfiniLM-suffixed InfiniOps references.
• Native-provider tests passed: kv_caching 36/36, random_sample 32/32, rope 162/162, paged_attention_prefill 76/76, and paged_attention 22/22.