[XLS JIT] Improve LLVM IR codegen for partitions, temporaries, literals, and selects

[copybara-service]

[XLS JIT] Improve LLVM IR codegen for partitions, temporaries, literals, and selects

Improve XLS JIT/AOT LLVM IR lowering in //xls/jit so that LLVM can inline and discard helper partitions, promote cross-partition temporaries to SSA registers, place constant literal tables in .rodata, and lower bit-level selects and boolean trees without control-flow branches or byte-padding masks:

1. Private partition functions and stack-promoted cross-partition temporaries: Mark partition functions (__<fn>_partition_<i>) with PrivateLinkage so LLVM discards standalone partition bodies once inlined into the entry function. Emit typed inbounds GEPs in GetOffsetIntoTempBuffer (instead of ptrtoint/add/inttoptr) and allocate a local stack alloca (local_temp_buffer, capped at kMaxLocalTempBufferBytes = 32 KiB) when a function has no early-exit/resume points (such as blocking proc receives/sends that require temp_buffer state to persist across invocations). This allows LLVM's SROA and dead-store elimination to promote cross-partition temporaries directly to SSA registers.
2. Read-only module-level constant globals for aggregate literals: Materialize all Literal nodes at their uses (ShouldMaterializeAtUse) and emit private unnamed_addr constant globals (CodeModel::Small) in NodeIrContext::GetOperandPtr (as well as on trace/assert string globals, with Cold on assertion failure calls) instead of re-initializing stack alloca buffers at runtime on every call.
3. SSA select lowering and i1 balanced reduction trees: Lower Bits-typed sel, priority_sel, and one_hot_sel directly in SSA registers (CreateSelect/CreateAnd/CreateOr) instead of emitting per-case basic-block branches (OrInPlace) or pointer-level select + memcpy. Truncate 1-bit (bits[1]) values in bit_slice, not, concat, n-ary bitwise ops, and one_hot_sel to i1 in SSA registers and reduce concat, n-ary ops, and one_hot_sel with a balanced binary tree (TreeReduce) so expression depth stays $O(\log N)$ within LLVM's SimplifyDemandedBits recursion limit (MaxAnalysisRecursionDepth = 6).

Runtime and Memory Performance Impact:

• CPU Runtime: Strictly faster for JIT and AOT-compiled XLS functions: eliminates per-invocation stack initialization of constant lookup tables, removes per-bit conditional branch mispredictions in one_hot_sel, folds 1-bit one_hot_sel chains into straight-line bitwise and/or instructions instead of cmovs, and eliminates memory round-trips through temp_buffer.
• Memory Usage: Zero heap overhead; cross-partition temporaries are promoted to registers (bounded by 32 KiB stack alloca prior to SROA), and aggregate literal tables are deduplicated in .rodata.