"dynamic memory allocation" may not have been the right term to use to denote moving beyond finite state machines, for an audience which wants to consider a Turing machine as having statically allocated infinite memory. But it makes sense if you think of a Turing machine's memory as only storing enough data for the portion of the tape which has already been accessed, and allocating new memory as new tape regions are needed.
It's true that the actual machine sitting on your desk only has a fixed finite memory. In that sense, without a peripheral unbounded store, it's not properly Turing complete.
It's also true that we often find it convenient to analyze it as Turing complete anyway, but this abstraction basically goes hand-in-hand with abstracting away its memory limitations. I don't see any point in pretending you can be Turing complete using only programs with a statically given memory bound (which amounts to a logarithmic bound on the number of states in a corresponding finite state machine); if we're going to pretend, let's pretend the memory is effectively infinite as well.
>>we often find it convenient to analyze it as Turing complete anyway
It's convenient, because the proofs will still be true even if memory sizes grow by 500x. Turing machines clearly don't have a real, material existence, similarly to real numbers (assuming that all actual numbers are finite). Real numbers can be considered as modeling integers that are arbitrarily large.
Even in a classic Turing machine I interpret the tape as a single pre-allocated array (of infinite size).
When talking about Turing-complete programming languages, a relaxed definition with finite-size memory is assumed, and to me it seems that it's just hair splitting whether program hits that limit via malloc() call or by advancing a pointer in an internal fixed-size array. Dynamic memory allocation is just a name for a bunch of bookkeeping done on a fixed-size memory.
In practice asm.js programs are not allowed to dynamically allocate any memory from the browser: a fixed-size array is created before the asm.js program starts. From the point of the VM programs are just shuffling bytes within a fixed-size array, and yet, we can run "Turing-complete" programs in asm.js.