optimize path input decoding to avoid allocations

[fwbrasil]

Problem

The decoding of inputs from paths is a hot operation in my bechmark implementation and has a high allocation rate of Tuple2 instances.

Solution

• Since decodedPathInputs follows the same order and has the same size as pathInputs at the end of a successful decoding, use an Array to collect the decoded inputs and positionally refer to the original path input using pathInputs when folding the results.
• Modify both matchPathInner and foldDecodedPathInputs to use index-based iteration.
• Avoid Vector.headAndTail calls since it produces a Tuple2 allocation. The JIT should have an easier time eliding this allocation but I think it's best to avoid it given how hot the method is.

Notes

• This is part of Reduce allocation rate #3552

[kciesielski]

LGTM, thanks!