deps: bump pliron to master (no_std/hashbrown) + AttributeDict encapsulation sweep

[nihalpasham]

deps: bump pliron to master af1b7d0 + AttributeDict encapsulation sweep

Tracks upstream pliron master at af1b7d07dda9, picking up:

• pliron #114: no_std support; pliron's maps move to hashbrown
• pliron fix(mir-importer): preserve runtime Index projections in Rvalue::Ref #121: graph traversals changed from recursive to iterative
• pliron feat(codegen): fma contraction and an opt -O3 pass to match nvcc defaults #117: simplify_cfg optimizer added to the llvm-opt pipeline

Both the root workspace and the isolated rustc-codegen-cuda workspace move to the same rev (required, or pliron resolves to two distinct crates and types stop unifying). Both lockfiles updated together.

Why the bump broke 4 call sites

hashbrown's get<Q: Hash + Equivalent<K>> no longer pins the key type, so call sites that reached into AttributeDict's inner map through the public .0 field with "literal".try_into().unwrap() keys stopped inferring (E0283: type annotations needed). Exactly 4 sites broke, all in mir-lower/src/convert/ops/memory.rs.

Call sites that already used the wrapper API (dialect-nvvm debug ops, AlignmentAttr helpers in llvm-export) were untouched by the change. The break only bit code that bypassed the abstraction.

The encapsulation sweep (the actual point)

Rather than patching the 4 sites with turbofish, this PR routes every keyed AttributeDict access through the typed wrapper, so no future map-internals change upstream can reach us:

Before: op.attributes.0.insert(key, attr.into())          // manual AttrObj boxing
After:  op.attributes.set(key, attr)                      // typed

Before: op.attributes.0.get(&key) + downcast_ref::<T>()   // two-step, two error paths
After:  op.attributes.get::<T>(&key)                      // downcasts for you

Before: op.attributes.0.contains_key(&gpu_kernel)
After:  op.attributes.get::<StringAttr>(&gpu_kernel).is_some()

Swept sites: callee/mutable stamps in the mir-importer translator, gpu_kernel + cluster-dim + launch-bound stamps and propagation (mir-importer body.rs, mir-lower lowering.rs), fastmath/nneg/overflow flag stamps (mir-lower arithmetic/cast/call), the nneg read and kernel metadata reads in the llvm-export emitter.

Whole-map operations keep using the public field because the wrapper has no equivalent yet: the MirFuncOp printer in dialect-mir (retain + is_empty on a clone_skip_outlined copy) and a keys() introspection in llvm-export's ops_test. Upstream wrapper candidates: remove/retain and key iteration.

One additional fallout fix: upstream relaxed set_operand_segment_sizes to take &Context; a mir-lower test passed &mut ctx and tripped clippy::unnecessary_mut_passed under -D warnings.

Verification

• 4 inference breaks reproduced pre-fix, workspace check clean post-fix
• All 5 compiler-crate test suites pass (101 tests, 0 failed); workspace test binaries compile
• fmt, clippy -D warnings (--all-targets), doc -D warnings, doctests: clean
• Backend rebuilt against the new rev; compile-only builds verified: vecadd (sm_90), tcgen05 (sm_100, 102 tcgen05 instructions in PTX), wgmma (sm_90a)
• Rewritten attribute paths verified end-to-end in artifacts: cluster emits cluster_dim_x/y/z nvvm.annotations; debug emits maxntidx/minctasm metadata and .maxntid in PTX

Labels: build-related, refactor, IR-lowering, llvm-export

GPU verification (done)

Full smoketest on a 2x RTX 5090 box (sm_120): compile-only 92/92, execution 91/92. The one execution failure is cluster Test 2 (cluster sync), which fails identically on current main on this hardware and is expected for now per the maintainer; not related to this bump.