[REFACTOR][IR] Unify Call construction and TVMScript printing

[tqchen]

Move best-effort result inference into the core Call constructor. Omitted or None ty uses available inference, explicit types are preserved, and Call.validate() checks the operator contract separately. I.Call directly exposes this constructor and preserves provisional IR without a separate unchecked path.

Use one shared printer for canonical named operations with operands, semantic attributes, type arguments and result types. Remove standard-call markers and dialect-wide registration helpers. Retain concrete hooks for representation-changing syntax, including LLVM names, PTX operands and Relax call_tir sugar; use exact raw I.Call reconstruction when named syntax cannot preserve the stored fields.

Preserve existing Python operator definitions, documentation and operand normalization, keeping helper parameters specific to their semantics and generating only missing exposures. Give assertions the stored condition/format/variadic-values contract and take-backward dedicated axis attributes, removing their special printer paths. Preserve contextual, native-call and pointer inference, reflect literal attribute defaults, and give the five legacy CUDA helpers explicit result-type parameters.