Using define method instead of class eval when defining model callbacks.

[wangjohn]

Based on my benchmark results, this change improves performance
substantially when defining callbacks (especially for apps that are defining large numbers of callbacks).

This benchmark (https://gist.github.com/wangjohn/5542610) was run using
the current master and also using my experimental branch which replaced
class_eval with define_single_method.

Using class_eval (current master):

                    user     system      total        real
10 trials       0.000000   0.000000   0.000000 (  0.001568)
50 trials       0.020000   0.000000   0.020000 (  0.021715)
500 trials      0.110000   0.000000   0.110000 (  0.115357)
1000 trials     0.250000   0.000000   0.250000 (  0.260025)
10000 trials    2.560000   0.000000   2.560000 (  2.568408)
50000 trials   12.800000   0.010000  12.810000 ( 12.886871)

Using define_single_method (experimental branch):

                    user     system      total        real
10 trials       0.000000   0.000000   0.000000 (  0.000790)
50 trials       0.000000   0.000000   0.000000 (  0.002960)
500 trials      0.050000   0.010000   0.060000 (  0.055690)
1000 trials     0.100000   0.000000   0.100000 (  0.094073)
10000 trials    0.890000   0.000000   0.890000 (  0.900364)
50000 trials    4.650000   0.000000   4.650000 (  4.686127)