Cache uniform locations in ffglex::FFGLShader

[vjandrea]

FFGLShader::FindUniform called glGetUniformLocation unconditionally on every call. Every Set(...)
overload routes through it, so a plugin setting a uniform every frame (e.g. Particles, which calls
FindUniform by name around 19 times per frame across two shaders) paid a string-keyed driver round-trip per
uniform per frame, per instance, for no benefit: a uniform's location is fixed for the lifetime of a linked
program.

Adds a name → location cache to FFGLShader, populated lazily in FindUniform and invalidated (cleared)
whenever the underlying program changes: in FreeGLResources(), and defensively at the top of
LinkProgram(). No public API change.

Fixes #80.

Notes for reviewers:

• Unbounded cache growth is intentional, not an oversight. uniformLocations never evicts entries. This
  is OK in practice because it's bounded by the number of distinct uniform names a given shader declares
  (single digits up to Particles' ~19, the largest user in this tree) not by frame count or anything
  that grows over the plugin's lifetime. An eviction policy would add complexity for a cache that already
  self-limits at the shader's actual uniform count.
• The cache-clear I added at the top of LinkProgram() is currently dead code. I traced every caller of
  Compile()/FreeGLResources() in this repo (Add, AddSubtract, Gradients, CustomThumbnail,
  Particles' GLResources, ffglquickstart::Plugin) and all of them follow the same
  InitGL → Compile() then DeInitGL → FreeGLResources() lifecycle, once each, per instance, and none of them
  call Compile() a second time on an already-linked shader without freeing it first. So today, clearing in
  FreeGLResources() alone would be enough to keep the cache correct. I kept the extra clear in
  LinkProgram() anyway as cheap insurance: nothing in FFGLShader's public API actually prevents a future
  caller from calling Compile() twice without an intervening FreeGLResources(), and if that ever
  happened, the old cache entries would silently point at locations from the previous (now-replaced)
  programID without this clear. Flagging so it doesn't read as accidental defensive-programming clutter because
  it's addressing a real (if currently unexercised) gap in the class's invariants. Please don't hesitate to flag it for removal anyway, it if you think that it's just premature optimization.

[vjandrea]

@MennoVink This is the same request as #83, closed without merging even after the find() based fix in that thread was confirmed correct.
Since the code wasn't the objection there, curious whether the underlying one still stands here: that caching in FFGLShader "doesn't help achieve maximum performance nor make anything easier" (your words on #83), versus staying a thin wrapper and leaving caching to callers. If that stance holds, happy to close this and document the manual cache-in-InitGL pattern instead.