PERF: floatify

[jbrockmendel]

claude says this avoids an allocation, but I'd like to get @WillAyd's thoughts.

9-10% speedup in pd.to_numeric on floaty strings:

N = 200_000
rng = np.random.default_rng(0)

shorts = pd.array([f"{v:.4f}" for v in rng.uniform(0, 1e4, N)], dtype=object)
longs = pd.array([f"{v:.17g}" for v in rng.uniform(0, 1e15, N)], dtype=object)

%timeit pd.to_numeric(shorts)
26 ms ± 203 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)    # <- main
23.8 ms ± 284 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # <- PR

%timeit pd.to_numeric(longs)
29.5 ms ± 306 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # <- main
26.7 ms ± 527 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # <- PR

[WillAyd]

Nice find and makes sense

[WillAyd]

I think we should just change the function to accept a const char * instead of casting (separate PR is fine)