[MLIR] Add C API for Enzyme dialect (MLIRCAPIEnzyme)

[agarret7]

Add C API for Enzyme MLIR dialect (MLIRCAPIEnzyme)

The Enzyme MLIR dialect has no C API, making it inaccessible to
foreign-language bindings (Rust, Python, etc.) without depending on C++
headers. This PR adds the missing layer, following the same pattern used
by upstream MLIR's own dialect C APIs (mlir-c/Dialect/Arith.h, etc.).

What's added

• Dialect registration — MLIR_DECLARE/DEFINE_CAPI_DIALECT_REGISTRATION
  so the dialect can be loaded into an MlirContext from C
• enzymeActivityAttrGet — constructs an enzyme.activity attribute
  from a C-side EnzymeActivity enum
• enzymeAutoDiffOpCreate / enzymeForwardDiffOpCreate — construct
  unattached enzyme.autodiff / enzyme.fwddiff ops from C, taking the
  same arguments as the TableGen-generated C++ builders
• Build system — Integrations/c/CMakeLists.txt with an
  add_mlir_public_c_api_library(MLIRCAPIEnzyme ...) target, wired into
  Integrations/CMakeLists.txt and MLIR/CMakeLists.txt

Not included (deferred)

enzymeRegisterPasses is not included because MLIREnzymeTransforms
depends on MLIREnzymeAnalysis, which currently fails to build due to a
pre-existing MLIR API change (getSuccessorInputs →
getSuccessorRegions), tracked in #2858.

Testing

MLIRCAPIEnzyme builds cleanly against LLVM/MLIR 22. The round-trip test will pass once the MLIREnzymeAnalysis build failures are resolved.

Edit: The test/MLIR/CAPI/roundtrip.mlir I included is a round-trip test for the ops themselves, not the C API. The existing ForwardMode/ReverseMode tests already cover that. A proper C API test requires a compiled C binary and lit infrastructure that doesn't exist in Enzyme's test suite yet. I've removed the placeholder. Happy to add proper C test infrastructure as a follow-up if that's the preferred direction.

[mofeing]

lgtm

[wsmoses]

@mofeing note once this lands, we should adapt reactantextra to remove its copy of similar apis

[wsmoses]

@agarret7 minor, but clang-format fails

[agarret7]

Updated this PR with the following additions:

• MLIR_CAPI_EXPORTED on all function definitions
• Pass creation functions: enzymeCreateBatchPass, enzymeCreateBatchDiffPass, enzymeCreateConvertEnzymeToMemRefPass, enzymeCreateRemoveUnusedEnzymeOpsPass,
  enzymeBatchOpCreate constructor
• Rebased onto current main

Motivation: I'm planning a PR to Melior to expose Enzyme-backed autodiff. Melior upgraded to MLIR 22 a couple months ago, so I validated this C API end-to-end against MLIR 22 via a small Rust test harness using bindgen.

[wsmoses]

sorry one more nit for the activity one, then lgtm

[wsmoses]

tangent [and @mofeing feel free to offer suggestions]

but can you also add this to the bazel build system too?

[agarret7]

tangent [and @mofeing feel free to offer suggestions]

but can you also add this to the bazel build system too?

I'm not super familiar with bazel, what would that entail?

[agarret7]

tangent [and @mofeing feel free to offer suggestions]
but can you also add this to the bazel build system too?

I'm not super familiar with bazel, what would that entail?

Looking at enzyme/BUILD, the CInterface target (BUILD:1052-1053) already builds Integrations/c/EnzymeMLIR.{cpp,h}. Maybe no changes are needed for to add to bazel? I don't have it set up locally to confirm with a build, happy to if someone can point me at what is needed to test it.