Here is a useful inequality for the operator norm of a block partitioned positive semi-definite matrix. This comes from Section 8 of Halko, Martinsson and Tropp's excellent writeup, which I recommend reading at least Sections 8 and 9 if you are interested in matrix analysis. $ \newcommand{\abs}[1]{| #1 |} \newcommand{\ind}{\mathbf{1}} \newcommand{\norm}[1]{\lVert #1 \rVert} \newcommand{\R}{\mathbb{R}} \newcommand{\ip}[2]{\langle #1, #2 \rangle} \newcommand{\T}{\mathsf{T}} \newcommand{\Proj}{\Pi} \newcommand{\Tr}{\mathrm{Tr}} $
Below, $\norm{\cdot}$ will refer to spectral (operator) norm of a matrix, and $\succcurlyeq$ refers to the PSD ordering
Proposition: Let $M = \begin{bmatrix} A & B \\ B^\T & C \end{bmatrix} \succcurlyeq 0$ be a block partitioned positive semi-definite matrix. We have that $$ \norm{M} \leq \norm{A} + \norm{C} \:. $$
Remark: Before we prove this proposition, we note that it immediately applies to block partitioned PSD matrices of the form $$ M = \begin{bmatrix} M_{11} & M_{12} & ... & M_{1k} \\ M_{12}^\T & M_{22} & ... & M_{2k} \\ \vdots & \vdots & \ddots & \vdots \\ M_{1k}^\T & ... & ... & M_{kk} \end{bmatrix} \:, $$ by recursively applying the proposition to the sub-blocks. That is, $$ \norm{M} \leq \sum_{i=1}^{k} \norm{M_{ii}} \:. $$ When the blocks are 1-by-1, this inequality reduces to $\norm{M} \leq \Tr(A)$, which is a crude bound, but sometimes useful.
Proof: We first need to prove a useful inequality, which is
$$
\begin{equation}
\norm{B}^2 \leq \norm{A} \norm{C} \:. \label{eq:hadamard_psd}
\end{equation}
$$
We first assume that $M \succ 0$, in which case both $A \succ 0$ and $C \succ 0$.
By Schur complements, we have that
$$
A \succcurlyeq B^\T C^{-1} B \succcurlyeq \frac{1}{\norm{C}} B^\T B \Longrightarrow \norm{B}^2 = \norm{B^\T B} \leq \norm{A}\norm{C} \:,
$$
which proves the inequality $\eqref{eq:hadamard_psd}$ when $M \succ 0$.
Now we assume $M \succcurlyeq 0$ only. Fix any $\varepsilon > 0$.
The matrix $M + \varepsilon I \succ 0$, for which we apply the inequality to conclude
$$
\norm{B^\T B} \leq \norm{A + \varepsilon I} \norm{C + \varepsilon I} = (\norm{A} + \varepsilon)(\norm{C} + \varepsilon) \:.
$$
Taking the limit as $\varepsilon$ tends to zero on the RHS yields the desired
inequality $\eqref{eq:hadamard_psd}$ when $M \succcurlyeq 0$.
Now for any vector $z = \begin{bmatrix} x \\ y \end{bmatrix}$ with $\norm{z} = 1$
and $x, y$ partitioned to conform with $M$,
$$
\begin{align*}
z^\T M z &= \begin{bmatrix} x \\ y \end{bmatrix}^\T \begin{bmatrix} A & B \\ B^\T & C \end{bmatrix} \begin{bmatrix} x \\ y \end{bmatrix} \\
&= x^\T A x + 2 x^\T B y + y^\T C y \\
&\leq \norm{A} \norm{x}^2 + \norm{C} \norm{y}^2 + 2\norm{B} \norm{x}\norm{y} \\
&\stackrel{(a)}{\leq} \norm{A} \norm{x}^2 + \norm{C} \norm{y}^2 + 2\norm{A}^{1/2}\norm{x} \cdot \norm{C}^{1/2} \norm{y} \\
&= (\norm{A}^{1/2} \norm{x} + \norm{C}^{1/2} \norm{y})^2 \\
&\stackrel{(b)}{\leq} (\norm{A} +\norm{C})(\norm{x}^2 + \norm{y}^2) \\
&= \norm{A} + \norm{C} \:,
\end{align*}
$$
where (a) uses the inequality $\eqref{eq:hadamard_psd}$, and (b) uses the Cauchy-Schwarz inequality.
Combining this calculation with the variation form of the operator norm,
$$
\norm{M} = \sup_{z = (x, y) : \norm{x}^2 + \norm{y}^2 = 1} z^\T M z \leq \norm{A} + \norm{C} \:,
$$
which is the desired inequality. $\square$