…rameter plumbing

Add infrastructure for prologue/epilogue fusion in generated Triton code:
- store_transform/load_transform hooks in memory_ops for hl.store/hl.load
- Extra params, removed args, and protected arg names in codegen pipeline
- dim_index_exprs on SubscriptIndexing for epilogue fusion offsets
- _ensure_inductor_fusion_config() to enable fusion via config_patches

stack-info: PR: #1724, branch: yf225/stack/71

# Conflicts:
#	helion/_compiler/generate_ast.py

…onTemplateBuffer

Add fusion code generation to template_buffer.py:
- _render_with_hooks: setup fusion hooks, read fusion metadata, build
  extra_params, pass store/load transforms conditionally
- _build_call_args: add prologue_source_buffers and extra_params support
- _generate_triton_ast: pass store/load transform and extra_params
- create(): add on_tensor_leaf/on_non_tensor_leaf callback params
- _codegen_epilogue_fusion: emit per-epilogue index definitions and
  STORE_OUTPUT placeholders
- _codegen_prologue_fusion: emit prologue variables and LOAD_INPUT
  placeholders
- _flatten_return_ast: helper for build_multi_outputs traversal
- lower_helion_kernel: compute epilogue_fusable_outputs from stored
  proxy ids and output fusion metadata

stack-info: PR: #1725, branch: yf225/stack/72

- Replace unconditional `self.skipTest` with `requires_torch_version("2.11")`
  guard, now that the required PyTorch-side changes have landed
- Update `expected_num_kernels` values across ~60 tests to reflect
  reduced kernel counts from prologue/epilogue fusion
- Un-skip 3 previously disabled tests:
  `test_autotune_no_fusion_final_has_fusion`,
  `test_inductor_output_code_has_helion_generated_triton_kernel`,
  `test_symint_return_from_tensor_shape`

stack-info: PR: #1727, branch: yf225/stack/74

The previous code from PR #1722 treated ALL unbacked symbols as
data-dependent, but compound expressions like `flag * 2` (= 2*u0)
have unbacked symbols that come from known kernel parameters.

Fix: check whether unbacked symbols are a subset of sym_remap keys.
If yes, substitute concrete values and evaluate. If any symbol is
unknown, raise DataDependentOutputShapeNotSupported.

Un-skip 3 tests that were broken by the overly broad check:
test_kernel_returns_none_in_tuple,
test_kernel_returns_none_first_in_tuple,
test_kernel_returns_tuple_of_scalars.

Add ref kernel equivalents and baseline kernel count verification to
test_torch_compile.py. Each test's `f` function now accepts a `_kernels`
kwarg, allowing `functools.partial` to swap helion kernels for eager
refs. The helper `_compile_and_count_kernels` is extracted from
`_run_compile_test`, which gains `ref_kernels` and
`expected_num_kernels_ref` parameters to assert that the ref
baseline produces the expected number of Triton kernels via
torch.compile/inductor.

…se 2)

When a Helion kernel's output feeds into a single-axis reduction
(e.g., k_rms_norm(x, w).sum(dim=1)), Inductor previously produced
2 kernels because the scheduler unconditionally blocked reduction
epilogues on templates. This change fuses supported reductions into
the template kernel, producing 1 kernel.

Phase 1 (persistent): When the tile covers the full reduction dim
(e.g., full slice ':' or block_size >= dim_size), the template emits
tl.sum/tl.max/tl.min inline at the store site.

Phase 2 (loop): When the tile is smaller than the reduction dim
(block_size < dim_size), the grid dim is converted to a for-loop
with an accumulator that sums partial reductions across iterations.

Only single-axis reductions with sum/max/min are supported.
Full reductions (sum()), argmax/argmin, welford, and multi-axis
reductions fall through to the 2-kernel path.