The master-slave arrangement doesn't strictly solve the metastability issue, AFAICT. It is commonly used to cross over between different clock domains of synchronous logic, but I don't quite see what improvement it does on purely asynchronous input (the slave gets a clear state, but it may be derived of a metastable transition anyway). It could simply be an incomplete description, as you could add a hysteresis function by combining the outputs of the two registers.
As for the differences between SR, JK, D or even T flip-flops, it tends to boil down to which inputs are asynchronous. The simplest SR latches do not toggle with S=R=1, but simply keep whichever state was kept last (or in the worst case, oscillate with a gate delay), that's the race. The JK, on the other hand, will transition on the clock edge - synchronous behaviour. It is thus their nature that a T register can only be synchronous, and an asynchronous D latch is transparent while latching. The SR register you describe doesn't have the T function, which can be useful depending on the function. For instance, a ripple counter can be described purely with T registers. Simply put, the JK gives you a complete set of operations (set, clear, toggle, and no-op) without costing an extra control line.
In synchronous logic, we frequently use wide sets of registers to implement a larger function. It doesn't strictly matter there if we use D, T, JK or whatever registers, as we can just redesign the logic function that drives them to include feedback (unless we need to build that logic - i.e. in 74 family logic). That's why FPGAs and such tend to have only D registers in their schematic representations. What does matter is that the register itself introduces the synchronous operation - steady state until the next clock. This allows combining plenty of side-by-side registers or ones with feedback functions.
As for the choice between delayed-pulse and clock-synchronous logic, it's not an automatic one. Some early computers (f.e. PDP-1) and even some highly energy efficient ones (f.e. GreenArrays) use the delayed-pulse design, and it is in fact comparable to a pipelined design in synchronous logic. The Carry-Save adder demonstrates the crucial difference - it's a pipelined design where you actually don't have a known value, not even intermediate, until the pulse from the last new value to enter has come out the other end. If you know at the logic design stage repeated accumulation but only the final sum is used, it may be the best choice. Meanwhile, FPGAs are typically designed with only a few clock nets and therefore do not adapt well to delayed-pulse logic (though it can be approximated with clock gating).
I hope this is more helpful than further confusing... interesting questions!
If you scrutinize the data sheet for a real D-FF carefully, you will see an item called 'setup time'. In actuality, the FF doesn't grab the value at the exact time of the clock edge; the data has to be stable for the last 20 ns or so before the clock rises, and that's the value that gets transferred. Also, the value at the output takes a few ns to settle down to the (possibly) changed value. So if you daisy chain a string of D-FF's together, Q from one into the D of the next, everything works because during the critical time for each stage's D input, the Q's are stable; the Q's only change very shortly after the active clock edge.
Best Answer
The input to D must not change for a minimum time after the clock. This is the hold time. It also must have stop changing a minimum about time before the clock. This is the setup time. If the input flip-flop is clocked slightly before the output flip-flop, the second D may be changing in the setup hold time window. This can cause erratic operation.