Problem

c_pointers() on CUTLASS DSL scalar types (Int32, Float32, Float64, Float16, BFloat16, TFloat32, etc.) creates fresh ctypes objects on every
kernel invocation. In ML training loops with frequent kernel launches, these short-lived allocations interleave with long-lived framework objects,
causing memory fragmentation and monotonic process RSS growth that eventually leads to OOM.

Observed in production: ~0.3 GB/step RSS growth in large model training, leading to OOM within hundreds of steps. MoE architectures (many expert
kernels per step) are particularly affected.

Root Cause

Each call to c_pointers() creates 3 ctypes objects (c_int/c_float, pointer(), cast()) that are immediately discarded after use. In a training
loop with multiple kernel calls per step, this produces hundreds to thousands of short-lived objects per step that fragment the heap. tracemalloc
cannot detect this because ctypes objects are freed before snapshots — the damage is heap fragmentation.

Fix

Add a per-type _cptr_cache dictionary that caches c_pointers() results keyed by scalar value. On subsequent calls with the same value, return
the cached result directly.

Related issue #3351