Implement fragment shader to support palette textures 'directly', and
therefore avoid having to reload all palette textures whenever the palette
changes.
src/pvr2/gl_sl.c
src/pvr2/glrender.c
src/pvr2/glutil.c
src/pvr2/glutil.h
src/pvr2/scene.c
src/pvr2/scene.h
src/pvr2/shaders.glsl
src/pvr2/texcache.c

Eliminate GL_REPLACE tex mode in favour of GL_MODULATE (by setting colour
values to 1.0) - one less case for shaders to care about later
src/pvr2/glrender.c
src/pvr2/scene.c

Bind texture 0 instead of enabling/disabling texturing all the time
(marginally faster, but also simpler)
src/pvr2/glrender.c
src/pvr2/rendsort.c

Change forced-depth-function lists (autosort, punchout) to just set the
depth once at the start of the list
Remove unused parameter from render_set_tsp_context
src/pvr2/glrender.c
src/pvr2/pvr2.h
src/pvr2/rendsort.c

Ensure stencil is disabled after use.
src/pvr2/glrender.c

Set gl texture parameters at texture load time rather than render time.
(This does mean that if the texture is used with variant parameters it will be
loaded multiple times). 3-4% faster this way
src/pvr2/glrender.c
src/pvr2/pvr2.h
src/pvr2/texcache.c

Dump more information with --gl-info, and print it a little more nicely
src/display.h
src/drivers/video_glx.c
src/drivers/video_nsgl.m
src/drivers/video_null.c
src/main.c
src/pvr2/gl_sl.c
src/pvr2/glutil.c
src/pvr2/glutil.h

Perform backface culling in scene preparation rather than leaving it to the
GL - this is a huge performance win, at least on the 9400M - changing cull
state appears to be very expensive, whereas the CPU needed to do the same
job is only just barely measurable.
src/pvr2/glrender.c
src/pvr2/rendsort.c
src/pvr2/scene.c
src/pvr2/scene.h

Turn on COLOR_SUM + FOG unconditionally, relying on scene setup to set
appropriate empty values when disabled - this is a minor performance win
src/pvr2/glrender.c
src/pvr2/scene.c

Add missing files for the randmath test, and update .hgignore
.hgignore
test/randmath.c
test/random.dat