Switch: turn on and distribute PGO profiles (follow-up to #87) #90

PR #87 adds profile-guided optimisation for the Switch build (-DWWHD_PGO=GEN|USE, tools/switch/pgo_train.py, docs/pgo.md), but it stays off by default and no release ships a PGO build for now. On a console at stock clocks it gave a further +2.6% to +7.5% fps on top of the recompiler changes (main thread -4.7% to -6.1%), measured at GPU 307 MHz (see docs/pgo-progress.md).

This issue is for turning it on for players later.

Why it is not on yet

• A profile only fits the exact sources it was made from. Those sources are each player's own recompile (from their own dump), so a profile depends on the recompiler version, the region (USA/EUR) and the runtime files it covers.
• GCC ignores the profile of any function whose flow graph changed (-Wno-coverage-mismatch), so a stale profile silently gives nothing, or only part of the gain.
• The .gcda files come from the game's code, so how they could be distributed (release asset, private pinned repo like the shader template in Ship a shader state manifest with releases so setup builds the shader cache before the first start #88, or not at all) needs the same care as any game-derived data.
• The GCC version used to build must match the one that made the profile.

To decide / do

• Whether a profile made by the maintainers on a reference recompile fits players' recompiles closely enough (same recompiler + same region should give identical generated code; check).
• How the profile is stored, pinned to a commit and fetched by build.sh / the installers (and checked against the toolchain version), or a clean fallback to a normal build.
• Per-region profiles or one shared profile.
• A release step that retrains after any change to the recompiler or a profiled runtime file (pgo_train.py, ~30 min on a console), in docs/release-checklist.md.
• Re-time PGO on current main at GPU 384 MHz (perf_test.py ab) to confirm the gain is still worth it.