[mlir][dxsa] Add gather4 variants

[asavonic]

Please ignore the first commit in the stack. It is reviewed separately in #182.

[tagolog]

With the latest PRs, I've introduced a new approach to test files that was approved by Anton.

Can we combine all gather-related tests to a single test file gather_ops.test and run mlir-translate with --split-input-file option?
atomic_ops.test

// RUN: mlir-translate --split-input-file --import-dxsa-hex %s | FileCheck %s
// RUN: mlir-translate --split-input-file --import-dxsa-hex %s | mlir-opt --split-input-file --verify-roundtrip

// CHECK-LABEL: dxsa.module {
// CHECK-NEXT:   dxsa.atomic_and u<0>, r<0>, r<1, <x>>
// CHECK-NEXT: }
0x070000a9, 0x0011e000, 0x00000000, 0x00100e46, 0x00000000, 0x0010000a, 0x00000001

// -----

// CHECK-LABEL: dxsa.module {
// CHECK-NEXT:   dxsa.atomic_and g<0>, r<2>, r<3, <x>>
// CHECK-NEXT: }
0x070000a9, 0x0011f000, 0x00000000, 0x00100e46, 0x00000002, 0x0010000a, 0x00000003

[asavonic]

What is the reason to split instructions into individual modules?

[tagolog]

A couple of reasons, most of them about making failures point at a single instruction rather than at a blob:

1. Per-variant round-trip. The --verify-roundtrip run is the main motivation. In one big dxsa.module the first round-trip mismatch aborts the whole check, so you never learn whether the remaining variants still round-trip. With --split-input-file each variant is verified independently, so a regression lands on exactly one block.

2. Independent import. If one opcode mis-decodes, it can shift the token stream for everything that follows in the same module and produce misleading cascade failures. Isolated modules keep a decode bug contained to its own case.

So the split is not about isolating the ops conceptually, it's about isolating failures per variant while keeping the whole family in a single reviewable file.

[asavonic]

Agree. Let's split the tests if we encounter any issues next time.