Program plan optimizations

[TristonianJones]

A significant amount of overhead comes from the instantiation of the dispatcher
with an immutable set of functions. Instead of recreating the dispatcher, create
a shared one across program instances.

Additionally, check whether optimizations are needed for an expression before
applying them as decorators.

Benchmark Case	Before (ns/op)	After (ns/op)	Δ Time	Before (B/op)	After (B/op)	Δ Memory	Before (allocs)	After (allocs)	Δ Allocs
BenchmarkProgramPlan/Default	8,153	930	-88.6%	8,320	1,416	-83.0%	36	26	-27.8%
BenchmarkProgramPlan/OptimizeUnneeded	7,344	1,150	-84.3%	8,784	1,512	-82.8%	50	32	-36.0%
BenchmarkProgramPlan/OptimizeNeeded	8,370	2,164	-74.1%	10,224	2,976	-70.9%	67	52	-22.4%

Addresses the dispatcher-sharing proposed in #1397, which is the first half of the optimizations suggested to minimize startup costs.

[TristonianJones]

@jpbetz FYI

[jpbetz]

LGTM, this looks like a cleaner/simpler approach than I had considered. Thanks!

[TristonianJones]

@jpbetz I went ahead and folded in more of the changes from your proposed PR. The numbers I see are close to what you observed:

pkg: github.com/google/cel-go/cel
cpu: Apple M1 Pro
BenchmarkEnvExtendK8sShape-10             	   30596	     38461 ns/op	   23332 B/op	     278 allocs/op
BenchmarkProgramK8sShape-10               	 1579915	       755.6 ns/op	    1144 B/op	      20 allocs/op
BenchmarkProgramFirstPerEnvK8sShape-10    	 1000000	      1440 ns/op	    1144 B/op	      20 allocs/op

There's two more changes to make to reduce the EnvExtend cost further, but then I think we'll have captured everything.

[jpbetz]

Awesome. This LGTM