I have read in many places that Fiat Shamir technique is used to convert constant round interactive protocols to non-interactive arguments. What would be the problem in applying Fiat Shamir to poly-round protocols?
Yes, there are examples of non-constant round interactive protocols that are unsound when the Fiat-Shamir transform is applied even in the random-oracle model. Note that for constant-round protocols soundness in the random oracle model was proved by Pointcheval and Stern [PS00] -- that is, any constant round interactive protocol (that has negligible soundness error) is also sound when the Fiat-Shamir transform is applied to it given the hash function is modelled as a random oracle.
One simple counter-example is to consider the sequential composition, say $n$ times, of a constant-round protocol that has constant soundness error. By amplification theorem, the resultant protocol has negligible soundness error --- inverse-exponential in $n$ to be precise --- but the Fiat-Shamir transformed protocol is totally insecure as argued next. Since the soundness error is constant a malicious prover can basically sample (in constant-many attempts in expectation) "favourable" messages for each round and with high probability come up with a cheating proof: see this lecture (around 11 minutes) by Ron Rothblum for more details.
[PS00] Pointcheval and Stern. Security arguments for digital signatures and blind signatures. Journal of Cryptology, 2000.