Mark primitives OPAQUE and update to clash-prelude-1.10

[felipebalbi]

Mark primitives OPAQUE on GHC ≥ 9.4, and support clash-prelude 1.10 / GHC 9.10

Two independent changes, one commit each, so you can take either on its own.

1. NOINLINE → OPAQUE (2d85eac)

Clash 1.10 warns on every primitive this package exports:

Warning: primitive Ice40.Rgb.rgbPrim isn't marked OPAQUE.
This might make Clash ignore this primitive.

clash-lib's isOpaque accepts only Opaque, so all twelve primitives trip it —
Spram, Spi, Rgb, Pll.Pad, Pll.Core, Osc (×2), Mac.Prim, Led, IO,
I2c, GB — not just the one a given design happens to use.

GHC 9.4 added OPAQUE for exactly this case. NOINLINE prevents inlining but still
lets the simplifier rewrite a binding in ways that can break blackbox matching;
OPAQUE also disables those transformations. For a primitive whose whole contract is
"match this name and emit the template", that is the stronger and more accurate
statement of intent.

It can't be silenced downstream. -fclash-no-prim-warn doesn't cover it —
clash-ghc emits this one through an unguarded warnIf in GenerateBindings.hs,
while opt_primWarn only guards "Dubious primitive instantiation" in clash-lib's
Netlist/BlackBox.hs. Confirmed by reading both sources and by running with the flag.

Why CPP. CI builds GHC 9.0, which has no OPAQUE, so each pragma is guarded and
9.0 keeps NOINLINE. That isn't a downgrade — on a GHC without OPAQUE there's
nothing stricter available.

2. Dependency bounds (6cdec17)

base < 4.20 → < 4.21, clash-prelude < 1.9 → < 1.11. Nothing in the package
actually breaks on the newer versions; the bounds were simply the binding constraint.

Granularity follows this package's own precedent: 1.6.1 was admitted with < 1.7 and
1.8 with < 1.9, so 1.10 gets < 1.11.

CI gains GHC 9.8 and 9.10 so the new bounds are tested rather than asserted — an
untested upper bound is how the previous one went stale, and I'd rather not repeat
that. 9.0 and 9.4 still resolve older clash-prelude, so the lower bounds are
untouched.

Verification

On GHC 9.10.3 / clash-prelude 1.10.0:

Check	Result
Library build, -Wall -Wcompat, all 13 modules	clean, zero warnings
Resolves without allow-newer	yes, after commit 2
Warning on a design instantiating SB_RGBA_DRV	gone
Generated Verilog, before vs after	byte-identical, md5 dc4bb4b391125dda03c4f6097b670c85

That last row is the important one: this changes what GHC is permitted to do, not
what Clash emits.

Notes

• Version left at 0.3.1.4, with the CHANGELOG entry under ## Unreleased — release
  timing is yours to decide. Happy to fold it into a 0.3.1.5 heading instead.
• CI runner is ubuntu-20.04, which GitHub has retired, so the workflow may not be
  starting at all today regardless of this PR. I left it alone as out of scope, but say
  the word and I'll bump it to ubuntu-latest here.
• cabal check reports missing upper bounds on ghc-typelits-{natnormalise,extra,knownnat}.
  Pre-existing, untouched.