std::vector spec leaves quite a bit of performance-critical details unspecified. I imagine the allocator behaviour differs enough between implementations that they implemented their own to keep performance in different platforms more predictable. The performance requirements for a physics engine can be quite different from the general case that the std::vector spec is aiming for.

I imagine templates are frowned upon for the same reason -- too many moving parts. Otherwise they're perfectly fine to use in projects that have goals other than portably predictable performance.