There are many instances where treating $dx$ as a number is convenient, and luckily, many times it can be at least partially justified. I'm sure your familiar with the 'proof' of the chain rule:
$$
\require{cancel}
\frac{dy}{du} \cdot \frac{du}{dx} = \frac{dy}{\cancel{du}} \cdot \frac{\cancel{du}}{dx} = \frac{dy}{dx} \, .
$$
While it is incorrect to naively view the chain rule as cancellation, I don't think this argument is that far off the truth. The chain rule can be more properly justified in the following way:
$$
\lim_{\Delta x \to 0}\frac{\Delta y}{\Delta x} = \lim_{\Delta x \to 0}\frac{\Delta y}{\Delta u} \cdot \frac{\Delta u}{\Delta x} = \lim_{\Delta x \to 0}\frac{\Delta y}{\Delta u} \cdot \lim_{\Delta x \to 0}\frac{\Delta u}{\Delta x} \, .
$$
It can then be shown that
$$
\lim_{\Delta x \to 0}\frac{\Delta y}{\Delta u} = \lim_{\Delta u \to 0}\frac{\Delta y}{\Delta u} \, .
$$
While even this argument sweeps some issues under the rug, the point remains that is often convenient to regard derivatives as fractions. And it shouldn't be too surprising that limits of fractions often behave just like fractions. It is clear that
$$
\lim_{h \to 0}\frac{f(x+h)-f(x)}{h} \approx \frac{f(x+h)-f(x)}{h}
$$
for small enough $h$. This means that $dy/dx$ can, roughly speaking, be viewed as the quotient of two tiny quantities, $dy$ and $dx$, which partially justifies the manipulation of them in an equation. There have been many attempts to formalise these heuristic arguments—for example, in nonstandard analysis. While I think most mathematicians will agree that none of these attempts have been entirely successful, it is reassuring to know that there is a framework that can turn these informal arguments into rigorous proofs.
$dy/dx$ can be viewed as an actual quotient in another sense, too. Say you plot the graph of $y=x^2$ and consider the gradient at a particular point $P$. If you move a certain distance along the tangent to $P$, going $dy$ units upwards and $dx$ units to the right, then $dy/dx$ should be exactly equal to $2x$. But if you 'zoom in' far enough, then moving along the tangent line looks no different to moving along the curve. This way of viewing $dy/dx$ as a quotient avoids the machinery of nonstandard analysis, while also being fairly intuitive in my opinion. This means that
$$
\frac{dy}{dx}=f'(x) \implies dy = f'(x)dx = \frac{dy}{dx}dx
$$
as you mentioned in your post. So $dx$ and $dy$ can be given a meaning as independent quantities, even though $dy/dx$ generally means 'apply the derivative operator, $\frac{d}{dx}$, to $y$'.
Regarding some of the other instances where you question the use of infinitesimals:
Integration by substitution
Integration by substitution comes from reversing the chain rule. Recall that if $y=f(g(x))$, then
$$
\frac{dy}{dx}=f'(g(x))g'(x) \, .
$$
Thus,
$$
\int f'(g(x))g'(x) \, dx = f(g(x))+C \, .
$$
On the other hand, if we make the substitutions $u=g(x)$ and $du = g'(x) \, dx$, then the integral becomes
$$
\int f'(u) \, du=f(u)+C=f(g(x))+C \, .
$$
It just so happens that working out $du/dx$ and then 'multiplying by $du$' means that you will make the correct substitutions. The formula can be summarised as
$$
\int f(g(x))g'(x) \, dx = \int f'(u) \, du \quad \text{where $u=g(x)$}
$$
In Leibnizian notation, it reads particularly well
$$
\int \frac{dy}{du} \cdot \frac{du}{dx} \, dx = \int \frac{dy}{du} \, du \, .
$$
Again, the $dx$'s 'cancel' in a similar fashion to the chain rule.
Separating the variables
It is common to solve first order differential equations in the following way:
\begin{align}
\frac{dy}{dx} &= f(x)g(y) \\[4pt]
dy &= f(x)g(y)dx \\[4pt]
\frac{1}{g(y)} dy &= f(x) \, dx \\[4pt]
\int \frac{1}{g(y)} dy &= \int f(x) \, dx \\[4pt]
\end{align}
Although the steps we used do seem questionable, the conclusion we draw is not. It can be directly shown that
$$
\frac{dy}{dx} = f(x)g(y) \iff \int \frac{1}{g(y)} dy = \int f(x) \, dx \, .
$$
Here is how:
\begin{align}
&\int \frac{1}{g(y)} dy = \int f(x) \, dx \\[4pt]
\iff & \frac{d}{dx} \int \frac{1}{g(y)} dy = \frac{d}{dx} \int f(x) \, dx \\[4pt]
\iff & \frac{1}{g(y)} \cdot \frac{dy}{dx} = f(x) \\[4pt]
\iff & \frac{dy}{dx} = f(x)g(y) \, .
\end{align}
I'm not sure if there is a 'deeper' reason why $dx$ and $dy$ can be treated as manipulable quantities in this case. Perhaps others can shed some light on that.
