perf(compute_mode): drop data.table overhead

[m-muecke]

Benchmarks via bench::mark (2000 iterations, ties_method = "first"):

input	old median	new median	speedup
int (n=100)	412µs	9.7µs	~42×
int (n=10k)	623µs	128µs	~4.9×
int (n=100k)	1.62ms	1.17ms	~1.4×
chr (n=100)	423µs	9.0µs	~47×
chr (n=10k)	690µs	258µs	~2.7×
dbl (n=100k)	2.76ms	1.87ms	~1.5×