Move Gradient inner loops into procs (except for perspective)

Mostly a refactoring, but a couple of substantive changes:

- replace some hardcoded 8s with appropriate symbolic constants (kCache32Shift)
- 16b TOGGLE_MASK was using the 32b cache size (old line 1010); switched to the 16b cache size (new line 142), which happens to be the same number at the moment.

We pass two float args to radial procs whether or not we're using float square roots; I'm open the alternative of changing the signature of the proc depending on that #define, but it'll be uglier.

[TomH, 2012-01-06 20:53:03]

This version passes GMs, and is currently being subjected to LayoutTests on ToT.

[TomH, 2012-01-06 21:49:41]

Passes LayoutTests with no new breakage.
Benchmarking is mostly the same, some improvement for mirrored gradients.

Fixed trailing-( style issues, added a missing semicolon that DEBUG build
caught.

[TomH, 2012-01-09 13:46:01]

Committed in r2988 with verbal LGTM from reed@.