Third, since $\sf{L} \subseteq \sf{NC}^2$, is there an algorithm to convert any logspace algorithm into a parallel version?
It can be shown (Arora and Barak textbook) given a $t(n)$-time TM $M$, that an oblivious TM $M'$ (i.e. a TM whose head movement is independent of its input $x$) can construct a circuit $C_n$ to compute $M(x)$ where $|x| = n$.
The proof sketch is along the lines of having $M'$ simulate $M$ and defining "snapshots" of its state (i.e. head positions, symbols at heads) at each time-step $t_i$ (think of a computational log). Each step $t_i$ can be computed from $x$ and the state $t_{i-1}$. Because each snapshot involves only a constant-sized string, and there exist only a constant amount of strings of that size, the snapshot at $t_i$ can be computed by a constant-sized circuit.
If you compose the constant-sized circuits for each $t_i$ we have a circuit that computes $M(x)$. Using this fact, along with the restriction that the language of $M$ is in $\sf{L}$ we see that our circuit $C_n$ is by definition logspace-uniform, where uniformity just means that our circuits in our circuit family $\{C_n\}$ computing $M(x)$ all have the same algorithm. Not a custom-made algorithm for each circuit operating on input size $n$.
Again, from the definition of uniformity we see that circuits deciding any language in $\sf{L}$ must have a function $\text{size}(n)$ computable in $O(\log n).$ The circuit family $\sf{AC}^1$ has at most $O(\log n)$ depth.
Finally it can be shown that $\sf{AC}^1 \subseteq \sf{NC}^2$ giving the relation in question.
Fourth, it sounds like most people assume that $\sf{NC} \neq \sf{P}$ in the same way that $\sf{P} \neq \sf{NP}$. What is the intuition behind this?
Before we go further, let us define what $\sf{P}$-completeness means.
A language $L$ is $\sf{P}$-complete if $L \in \sf{P}$ and every language in $\sf{P}$ is logspace reducible to it. Additionally, if $L$ is $\sf{P}$-complete then the following are true
$L \in \sf{NC} \iff \sf{P} = \sf{NC}$
$L \in \sf{L} \iff \sf{P} = \sf{L}$
Now we consider $\sf{NC}$ to be the class of languages efficiently decided by a parallel computer (our circuit). There are some problems in $\sf{P}$ that seem to resist any attempt at parallelization (i.e. Linear Programming, and Circuit Value Problem). That is to say, certain problems require computation to be done in a step-wise fashion.
For example, the Circuit Value Problem is defined as:
Given a circuit $C$, input $x$, and a gate $g \in C$, what is the output of $g$ on $C(x)$?
We do not know how to compute this any better than computing all the gates $g'$ that come before $g$. Given some of them may be computed in parallel, for example if they all occur at some time-step $t_i$, but we dont know how compute the output of gates at timestep $t_i$ and time-step $t_{i+1}$ for the obvious difficulty that gates at $t_{i+1}$ require the output of gates at $t_i$!
This is the intuition behind $\sf{NC} \neq \sf{P}$.
Limits to Parallel Computation is a book about $\sf{P}$-Completeness in similar vein of Garey & Johnson's $\sf{NP}$-Completeness book.