24d10 is 24 independently rolled d10's, that means we convolute the uniform distribution with itself 24 times and the result will be nearly indistinguishable from a scaled bell curve.
Edit: Click the “Graph” button here to visualize: http://anydice.com/program/4ddf
For large dice counts like that, it's probably better to just use a manual table for a bell curve based on ~8-10 flips. No human player is going to care about result fidelity beyond a thousand possible results anyway.
That gets you to 4.48 amortized flips per d10.
Above 3d10 the rate of re-rolls increases; 20 flips for 6d10 gives a worse amortized cost than two sets of three.
This has the guaranteed correct distribution ('correct' being 'uniform'), too!
EDIT: I guess, to be more clear about the algorithm: flip 8 coins and convert this result to binary. If the result is greater than 240, flip the eight coins again. The number of total 8-coin-flips that need to be done on average is something like 1.062, and the likelihood that you need to perform n 8-coin-flips to receive a result ≤ 240 vanishes exponentially as (15/256)^(-n).
EDIT: I screwed up again, 24d10 is 24 dice with 10 values. Oops. Point is still roughly the same!
24d10 becomes 0-240, so change it to 0-256 and you've got 8 flips (8d2?)
Distribution won't be correct but hey, 8 flips