feat: improve recursive diff performance

[cn-stephen]

Benchmark	Mode	Master	recursive-perf	Δ
PandaCSS Config	no cycles	12.06 µs	10.62 µs	−11.9%
PandaCSS Config	cycles	17.48 µs	12.88 µs	−26.3%
Scalar/OpenAPI Doc	no cycles	7.44 µs	6.76 µs	−9.1%
Scalar/OpenAPI Doc	cycles	10.80 µs	8.14 µs	−24.6%
Vanilla Extract Theme	no cycles	6.99 µs	6.43 µs	−8.0%
Vanilla Extract Theme	cycles	9.07 µs	7.44 µs	−18.0%
Large Object (300 props)	no cycles	14.01 µs	13.62 µs	−2.8%
Large Object (300 props)	cycles	13.84 µs	13.76 µs	−0.6%
Large Diff (300 props)	no cycles	9.74 µs	9.42 µs	−3.3%
Large Diff (300 props)	cycles	9.62 µs	9.48 µs	−1.5%
Deeply Nested Object	no cycles	144.1 µs	120.1 µs	−16.7%
Deeply Nested Object	cycles	180.9 µs	131.4 µs	−27.3%
Small Object	no cycles	65.81 ns	67.78 ns	+3.0%
Small Object	cycles	65.37 ns	66.95 ns	+2.4%

[AsyncBanana]

Have you tried mutating an array instead of a set? I haven't tested it, but it might be faster because it will avoid all of the hash table's overhead and shouldn't require any searches for insertion/deletion (you should just be able to insert/delete from the end).

[cn-stephen]

YEP! very clever. Array wins!

Benchmark	Mode	Master	recursive-perf	Δ (array)	Δ (Set)
PandaCSS Config	no cycles	12.06 µs	10.51 µs	−12.9%	−11.9%
PandaCSS Config	cycles	17.48 µs	10.66 µs	−39.0%	−26.3%
Scalar/OpenAPI Doc	no cycles	7.44 µs	6.80 µs	−8.6%	−9.1%
Scalar/OpenAPI Doc	cycles	10.80 µs	6.93 µs	−35.8%	−24.6%
Vanilla Extract Theme	no cycles	6.99 µs	6.43 µs	−8.0%	−8.0%
Vanilla Extract Theme	cycles	9.07 µs	6.49 µs	−28.4%	−18.0%
Large Object (300 props)	no cycles	14.01 µs	13.61 µs	−2.9%	−2.8%
Large Object (300 props)	cycles	13.84 µs	13.69 µs	−1.1%	−0.6%
Large Diff (300 props)	no cycles	9.74 µs	9.61 µs	−1.3%	−3.3%
Large Diff (300 props)	cycles	9.62 µs	9.49 µs	−1.4%	−1.5%
Deeply Nested Object	no cycles	144.1 µs	126.2 µs	−12.4%	−16.7%
Deeply Nested Object	cycles	180.9 µs	130.5 µs	−27.9%	−27.3%
Small Object	no cycles	65.81 ns	64.33 ns	−2.2%	+3.0%
Small Object	cycles	65.37 ns	62.75 ns	−4.0%	+2.4%

[cn-stephen]

I think we can remove cyclesFix option altogether

[AsyncBanana]

Those results are great! I am going to run a few more tests, but I think you might be right. The overhead is so negligible at this point that we might as well have it on by default.

[AsyncBanana]

I haven't been able to test everything yet, but I am going to merge this anyway. Thanks!