compiler: Do not CSE-capture IndexedPointer indices - #3047

Problem

IndexedPointer (ptr->field[idx]) had no handler in CSE's _catch, so the generic handler recursed into its index. When an integer index expression repeats inside a floating-point Cluster, CSE binds it to a CTemp, and cse_dtype promotes an integer expression in a float Cluster to the Cluster's float type. The result is a floating-point array subscript:

float r4 = (time / ct0f) + static_cast<int>(time >= 3)*static_cast<int>((3)%(ct0f) != 0);
f_vec->sizes[r4] = f_vec->sizes[r4];   // nvcc: expression must have integral or unscoped enum type
                                        // hipcc: array subscript is not an integer

This surfaced after #3043: the CondEq snapshot shift turned a bare IntDiv index (never captured) into an Add that CSE does capture. Any backend that lowers metadata lookups to IndexedPointer inside a float Cluster hits it.

Fix

Register IndexedPointer with the same _catch handler as Indexed/Symbol/Reserved, so CSE treats it as an opaque terminal and never looks inside its index. Two executable lines; no change to cse_dtype, whose promotion rule affects unrelated arithmetic.

Test

tests/test_cse.py::test_indexedpointer_index_not_captured: two IndexedPointers sharing a repeated integer index in a float32 Cluster; asserts no CTemp ends up inside an IndexedPointer index.

• Negative control: with devito/passes/clusters/cse.py reverted the test fails (assert not True: the index was captured); with the change it passes.
• tests/test_cse.py, test_dse.py, test_buffering.py: 306 passed, 2 xfailed (one MPI test fails identically without this change on the host used, due to a missing mpirun).
• Generated CUDA source compiled with nvcc 12.4: 5 errors before, 0 after; the same case runs on an RTX 3090 (CUDA 12.9).

🤖 Generated with Claude Code

CI note. pytest-osx-py312-clang-omp-base fails on this PR for a reason unrelated to the change: the runner it drew is a non-M1 Apple chip, which get_platform() does not classify as AppleArm, so test_nested_nthreads is not skipped and emits x86 intrinsics headers. Fixed in #3048. pytest-gpu-omp-amd segfaults in the examples step (test_tti_stability[staggered]); the generated code for that operator is byte-identical with and without this change, and a concurrent unrelated PR run failed the same job (runner contention, tracked in #3049).