Materialize isbits boxed constants as device globals

[maleadt]

Since JuliaLang/julia#55045 (1.14.0-DEV.1348), small isbits unions built from
constants are kept fully boxed: codegen emits a julia.constgv slot holding a
pointer to the host box, which device code then dereferences for the payload
or type tag. relocate_gvs! used to bake that host address into the IR, so any
such kernel reads host memory → ERROR_ILLEGAL_ADDRESS.

Fixing this upstream was rejected (JuliaLang/julia#62191): inline storage is a
host-side de-optimization, and flag-guarding every eager-box site doesn't
scale. Instead, this implements @topolarity's suggestion on the GPUCompiler
side, where the GC-related objections don't apply.

relocate_gvs! now inspects the object behind each constgv slot. Non-ghost,
non-Bool isbits objects are materialized: a private constant replica of
the box (header word + payload bytes, heap-compatible layout and alignment) is
emitted into the module and the slot points at its payload. Everything else —
types, symbols, singletons, Bools, mutable objects — keeps the baked address;
those are only ever used as identity tokens.

Notes:

• The replica's header holds the same value codegen compares against in
  typeof/isa (smalltag immediate or type pointer), and === on isbits
  falls back to bits comparison, so no identity-sensitive behavior changes.
• Resolution runs before optimize!, so payload and tag loads constant-fold;
  the device code ends up identical to what inline union storage produced
  before #55045. No de-optimization, host codegen untouched.
• Every eager-boxing site funnels through literal_pointer_val_slot, so
  future upstream boxing patterns are covered automatically.
• Session-dependence marking is now precise: relocate_gvs! reports whether
  any host address survived (baked slot, or a materialized header carrying a
  non-smalltag type pointer). Kernels whose constants are all smalltag
  (Int/UInt/Char etc.) no longer get wiped from package images.

Applies to 1.13+ (older versions bake addresses in Julia itself); the union
regression only exists on 1.14.0-DEV.1348+.

[codecov]

Codecov Report

❌ Patch coverage is 93.93939% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 81.57%. Comparing base (21b585d) to head (74386f6).
⚠️ Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
src/jlgen.jl	93.44%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #874      +/-   ##
==========================================
+ Coverage   81.35%   81.57%   +0.22%     
==========================================
  Files          25       25              
  Lines        4789     4847      +58     
==========================================
+ Hits         3896     3954      +58     
  Misses        893      893              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[topolarity]

Nice! This ended up even simpler than I hoped 👍

This seems like the better path to me. Something like JuliaLang/julia#62191 is possible, but it requires a more significant refactor to give any assurances that it's complete. This is complete by design, assuming it works, so that's quite nice.

Only concern is Bool, which might be worthy of more test coverage. Some notes inline.