Set booster param nthread only at first model loading.

[ericangelokim]

Description of changes:

The method load_model is called every invocation to get the cls.format of the booster. However, cls.booster.set_param('nthread', 1) was reset every time the method was called, and in XGB 0.90 this caused a latency difference in model loading of almost two magnitudes.

# Current
#       ContentTypeParsing  DataParsing  ModelLoading  Preprediction   Prediction  PostPrediction  TotalRuntimes
#count         5000.000000  5000.000000   5000.000000    5000.000000  5000.000000     5000.000000    5000.000000
#mean             0.000232     0.001062      0.000614       0.000016     0.000095        0.000026       0.002044
#std              0.000015     0.000086      0.000155       0.000001     0.000032        0.000003       0.000247
#min              0.000219     0.001000      0.000592       0.000015     0.000087        0.000024       0.001950
#50%              0.000227     0.001046      0.000601       0.000016     0.000090        0.000025       0.002010
#75%              0.000231     0.001074      0.000608       0.000016     0.000093        0.000026       0.002049
#90%              0.000250     0.001118      0.000643       0.000017     0.000106        0.000027       0.002137
#99%              0.000274     0.001264      0.000699       0.000018     0.000131        0.000036       0.002392
#max              0.000405     0.002798      0.011242       0.000049     0.001668        0.000071       0.016163

# Pushing setting nthreads = 1
#       ContentTypeParsing  DataParsing  ModelLoading  Preprediction   Prediction  PostPrediction  TotalRuntimes
#count         5000.000000  5000.000000   5000.000000    5000.000000  5000.000000     5000.000000    5000.000000
#mean             0.000228     0.001069      0.000008       0.000014     0.000089        0.000026       0.001435
#std              0.000016     0.000090      0.000063       0.000001     0.000028        0.000002       0.000175
#min              0.000216     0.001004      0.000004       0.000012     0.000081        0.000024       0.001348
#50%              0.000223     0.001054      0.000005       0.000014     0.000084        0.000026       0.001409
#75%              0.000227     0.001082      0.000006       0.000014     0.000087        0.000027       0.001444
#90%              0.000246     0.001119      0.000006       0.000015     0.000100        0.000027       0.001492
#99%              0.000271     0.001264      0.000007       0.000019     0.000128        0.000034       0.001707
#max              0.000465     0.002782      0.001994       0.000055     0.000908        0.000080       0.006040

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[aws-patlin]

Great work finding this bug!