Avoids using lambdas in MacroEvaluator hot paths.

[tgregg]

Description of changes:

These three lambdas were all in the top 9 contributors to total allocations when reading test data. After this change, they're gone. Note: the new allocation profile results show that the JVM optimizes away the no-op lambda.

Also note: beginning with this PR I've moved to version of the test dataset that does not use make_string invocations, as they are an optional feature that users may choose to use when data size is more important than read performance. As a result, these numbers are not directly comparable to the numbers in preceding PRs.

Speed: 219 ms/op -> 219 ms/op (0%)
Allocation rate: 124 KB/op -> 113 KB/op (-8.9%)

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[jobarr-amzn]

This doesn't seem like a good place to be in. A 9% drop in allocation is definitely not nothing, but the 0% change in speed makes me wonder how much allocation we have to improve before we're faster.

Can you add comments in these sites explaining why you're not using an initialization lambda, or a comment on _Private_RecyclingStack#push explaining that capturing lambdas ought to be avoided?