[WIP] difftests: integrate spirti for emulating SPIR-V shaders on the CPU.

[eddyb]

This PR uses spirti from the branch corresponding to this (yet unlanded) PR:

• [WIP] spirti: SPIR-T interpreter (on the CPU, optionally via rayon). spirt#46

---

I've been testing with this command:

SPIRTI_DEBUG=trace-all DIFFTEST_SPIRTI= cargo nextest run -P difftests -p difftests --release --no-default-features --features use-installed-tools --no-capture

The DIFFTEST_SPIRTI env var bypasses the wgpu runner of difftests, and feeds the SPIR-V blob through spirti instead (alongside with plumbing buffer inputs and outputs).

Just to get any tests passing, spirti was augmented with:

• non-uniform pointers (e.g. indexing arrays by the invocation ID)
• global-vs-local memory scoping (akin to semantic address spaces)
• non-uniform control-flow (a la "execution masks")

As a result, this is the summary from the above command:

     Summary [  71.626s] 21 tests run: 10 passed, 11 failed, 0 skipped
        FAIL [  21.707s] ( 1/21) difftests::difftests difftests::arch::atomic_ops
        FAIL [   1.706s] ( 2/21) difftests::difftests difftests::arch::memory_barriers
        FAIL [   2.757s] ( 4/21) difftests::difftests difftests::arch::vector_extract_insert
        FAIL [   1.741s] ( 5/21) difftests::difftests difftests::arch::workgroup_memory
        FAIL [   1.753s] ( 9/21) difftests::difftests difftests::lang::control_flow
        FAIL [   2.744s] (10/21) difftests::difftests difftests::lang::control_flow_complex
        FAIL [   2.797s] (13/21) difftests::difftests difftests::lang::core::ops::bitwise_ops
        FAIL [   3.773s] (14/21) difftests::difftests difftests::lang::core::ops::const_fold_int
        FAIL [   1.796s] (15/21) difftests::difftests difftests::lang::core::ops::math_ops
        FAIL [   1.772s] (17/21) difftests::difftests difftests::lang::core::ops::trig_ops
        FAIL [   1.782s] (19/21) difftests::difftests difftests::lang::core::ops::vector_swizzle

Some of the remaining failures seem potentially easy to fix, with workgroup-shared memory semantics being the one that's least obvious how to best approach.