The programs need to reflect the link state, which may be different
than the serialized state. This was done by creating temporary
shaders, link the program, delete the temporaries and attach the ones
from the state. This change simplifies this by:

 - creating the programs before the shaders
 - using the same shader ids for the temporaries as was serialized
 - attaching the shaders to the programs after creating the shaders

This has the benefit of the rebuilt state matching the serialized
state exactly (not just semantically) and no temporary shader IDs
have to be allocated.

`make<bool>(arena)` should return `false` (the zero value), but returns
`true` for any non-null arena, because the template expands to
`bool(arena)`, instead of `bool()` as pointers can be coerced to bools.
This adds the special no-arg case, forcing it to ignore the arena.

`eglMakeCurrent` resets the scissor and viewport box of the
context, but only does so the very first time the context is made
current. We currently were doing it every time the context is made
current. This change moves the detection of whether the reset is
needed from the extras into the API definition. It is impossible
to detect in the spy whether the context was just made current
for the very first time.

Note that the replay's `bindRenderer` now no longer will ever be
asked to reset the viewport because:

  1. For non-MEC replays, we can just rely on the OS's context
     `makeCurrent` call to do it for us (all of them do). This now
     works, because the replay no longer creates dummy surfaces.
  2. For MEC replays, the state reconstruction emits calls to set
     the viewport/scissor to the expected values.