refactor(codegen): unify review follow-ups

[zhen8838]

Why

• Close the review feedback on duplicated codegen traversal, launch derivation, inspection formatting wrappers, and ambiguous analysis work fields.
• Keep launch ABI and analysis counting domains explicit without adding parallel abstractions.

What

• Build one root CodegenContext, derive target-local emit contexts, and route common C++ plus CUDA-specific TIR traversal through CppEmitter and CudaEmitter visitors.
• Resolve launch edges and callable symbols before emission; represent the one-dimensional grid_x and block_x launch convention as scalar signature parameters.
• Replace compute-cost storage with generic Spread[V], rename analysis work service to other_ops, and render metadata directly in type-specific printer methods.
• Remove obsolete topology helpers, per-statement emitters, intermediate inspection values, and tests for deleted objects.

Contract

• compile() runs the default pass pipeline and then calls the lower-level build() path for pass-prepared TIR.
• flops and other_ops expose independent logical, total, and topology-ordered per_unit values while preserving dtype or operation-kind breakdowns.
• Target-side service throughput names remain unchanged.
• CPU-to-device launches use concrete host-computable scalar expressions; launch geometry has no None fallback.
• No new tests are added in this follow-up.

Risk

• Codegen currently supports one CUDA device target per linked artifact and one-dimensional launch topology.