chore(nccl): gate CGo path behind nccl_cgo build tag

Renames the legacy nccl.go to nccl_cgo.go and tightens its build tag to
//go:build cuda && cgo && nccl_cgo so the CGo implementation is OFF by
default. The new purego/dlopen binding in nccl_purego.go is the default
and only path on every supported platform; the CGo file is retained as a
debugging fallback.

Author: dndungu

internal/nccl/nccl.go internal/nccl/nccl_cgo.gointernal/nccl/nccl.go renamed to internal/nccl/nccl_cgo.go

@@ -1,9 +1,8 @@
-//go:build cuda
+//go:build cuda && cgo && nccl_cgo
 
-// Package nccl provides CGo bindings for the NVIDIA Collective Communications
-// Library (NCCL). It exposes communicator lifecycle, all-reduce, and broadcast
-// operations needed for multi-GPU gradient synchronization and tensor
-// distribution.
+// Legacy CGo binding for libnccl, retained as an opt-in fallback. Build with
+// `-tags "cuda nccl_cgo"` to use this implementation instead of the default
+// purego/dlopen path in nccl_purego.go.
 package nccl
 
 /*