Fix SM90 MegaMoE .template disambiguation on nvcc13

[Helicoplacoidea]

Summary

Pure compile-fix for SM90 FP8 MegaMoE on nvcc 13.x. Adds the standard dependent-name disambiguator .template to four chained …get_data_buffer(…).get_base_ptr<float>() call sites in sm90_fp8_mega_moe_impl().

Problem

On nvcc 13.x (EDG frontend), the chained call

(*l1_topk_weights_buffer).get_data_buffer(idx).get_base_ptr<float>()

fails to compile inside the deferred/dependent scope of the SM90 MegaMoE epilogue, with:

error: type name is not allowed
error: expected an expression

because EDG parses the unqualified < as the less-than operator instead of the start of an explicit template-argument list.

Fix

Add the standard dependent-name disambiguator .template to the four chained .get_data_buffer(…).get_base_ptr<float>() call sites reported by nvcc13 (L1655 / L1666 / L1824 / L1827):

-    .get_base_ptr<float>();
+    .template get_base_ptr<float>();

Compatibility

Behavior is unchanged on nvcc ≤ 12, where .template is a no-op when not strictly required.

Test Plan

• nvcc 13.x: JIT SM90 fp8 mega-moe kernel compiles.
• nvcc 12.x: regress to the same kernel, confirm no behavior change.