Salsa20 is a stream cipher based on a pseudorandom function, not a pseudorandom permutation.
For a fixed key $k$ and nonce $n$, the mapping $PRF^{S20}_{k,n}: \{0,1\}^{64} \to \{0,1\}^{512}$, which maps a "Stream position" to "keystream block", is supposed to be a pseudorandom function. It is not supposed to be injective (i.e. a permutation, even less since the input/output sizes are different), and there certainly is no easy way to produce a preimage, even knowing the key and nonce.
What you propose is the Salsa20 encryption function for the first "block" of plaintext:
$$\def\Enc{\operatorname{Enc}}\Enc^{S20}_{k, n} : \{0,1\}^{512} \to \{0,1\}^{512}$$
$$\Enc^{S20}_{k, n}(P) = P \oplus PRF^{S20}_{k,n}(0)$$
This is obviously bijective (it is its own inverse), i.e. a permutation, but at the same time, it is obviously not a (pseudo)random permutation. In addition to the self-inverse-property, it for example also has the two-time-pad property:
$$\Enc^{S20}_{k, n}(P_1) \oplus \Enc^{S20}_{k, n}(P_2) = P_1 \oplus P_2,$$
which no random permutation would have.