feat: caching code in mapreduce style ops

[avik-pal]

No description provided.

[codecov]

Codecov Report

❌ Patch coverage is 85.71429% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 68.86%. Comparing base (8040a3e) to head (8a5ef1d).

Files with missing lines	Patch %	Lines
src/TracedRArray.jl	85.71%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1662      +/-   ##
==========================================
+ Coverage   68.82%   68.86%   +0.04%     
==========================================
  Files         103      103              
  Lines       11596    11599       +3     
==========================================
+ Hits         7981     7988       +7     
+ Misses       3615     3611       -4     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[jumerckx]

Trying to understand what this is doing. Why isn't call_with_reactant required anymore, or why was it required before?

[avik-pal]

This is meant to be only used instead broadcast which with force the call_via_reactant in elem_apply (it should have not been added in the first place)

[jumerckx]

So this looks reasonable to me (though not sure if the CI failure is related to this pr?)

One major caveat is that the caching mechanism for Ops.call isn't correct if the function mutates a non-traced Julia object. e.g.:
for

function f(X, y)
	push!(X, y)
end
# X::Vector{TracedRArray}
# y::TracedRArray

because in this case tracing through f has a side-effect that isn't captured when simply inserting a func.call to the existing func.func.

Billy mentioned the need for an analysis that tells whether such a mutation occurs. I've looked into this a little bit but haven't yet found a good way to do this.

[avik-pal]

Ah I see, in that case, let's hold off on the caching part. I will separate out the non-caching fixes in a separate PR