Performance - class style fast path

[SanderMuller]

When class() or style() is called with a plain string (the common case), bypass the full merge() cycle and directly append the value to the attributes array. Avoids the escape loop, attribute partitioning, and array_merge overhead.

Biggest single win: -66% on the class benchmark (blaze/blade ratio).

[SanderMuller]

/benchmark class

[github-actions]

Benchmark Result: Default

Attempt	Blade	Blaze	Improvement
#1	343.05ms	15.77ms	95.4%
#2	343.58ms	15.64ms	95.4%
#3	342.02ms	15.97ms	95.3%
#4	347.83ms	15.90ms	95.4%
#5	342.64ms	15.85ms	95.4%
#6	351.02ms	15.81ms	95.5%
#7	349.20ms	16.07ms	95.4%
#8	341.42ms	15.95ms	95.3%
#9	347.39ms	15.83ms	95.4%
#10 *	345.19ms	16.25ms	95.3%
Snapshot	343.96ms	16.00ms	95.3%
Result	343.58ms (~)	15.85ms (~)	95.4% (~)

_{Median of 10 attempts (* = outlier, excluded from result), 5000 iterations x 10 rounds, 45.55s total}

_{To run a specific benchmark, comment /benchmark <name> where name is one of: attributes, aware, class, default, forwarding, merge, named-slots, no-attributes, slot}

[github-actions]

Benchmark Result: Class

Attempt	Blade	Blaze	Improvement
#1	239.29ms	13.67ms	94.3%
#2	238.79ms	13.49ms	94.4%
#3	234.62ms	13.21ms	94.4%
#4	237.78ms	13.33ms	94.4%
#5	239.65ms	13.38ms	94.4%
#6	233.81ms	13.52ms	94.2%
#7	235.35ms	13.42ms	94.3%
#8	236.71ms	13.83ms	94.2%
#9	231.54ms	13.69ms	94.1%
#10 *	232.39ms	14.17ms	93.9%
Snapshot	239.98ms	18.00ms	92.5%
Result	236.71ms (~)	13.49ms (-25.1%)	94.3% (+1.8%)

_{Median of 10 attempts (* = outlier, excluded from result), 5000 iterations x 10 rounds, 32.16s total}

_{To run a specific benchmark, comment /benchmark <name> where name is one of: attributes, aware, class, default, forwarding, merge, named-slots, no-attributes, slot}

[ganyicz]

I like this but feels like we're duplicating a lot of code also in #161. Would you be able to merge these together into one PR and see if we can extract some of the logic into separate methods please?