Split the gateway errors into two enums

[pierre-l]

See #282

Instead of relying on a single enum with a continuous error range, split the gateway errors into two enums:

1. IrrecoverableError, leave it a reasonable range (0..100?)
2. RecoverableError with the rest of the range (101..256?)

Add tests so that these indices don't change without a good reason (breaking change warnings)

[eloylp]

To take into account #238